Publication Record
Evidence current through: Thursday, November 19, 2026, at 12:45 a.m. UTC (7:45 p.m. ET and 4:45 p.m. PT on Wednesday, November 18).
Most of the infrastructure Microsoft announced at Ignite on November 17, 2026, is in preview, and the note reads it at the status Microsoft gave it by the time above. The sources were checked again for this edition, including Microsoft’s longer Ignite posts on Kubernetes and on confidential computing. Developments after the time above are not reflected. Times in the text are given in UTC.
Research Note In Two MinutesNote synopsis · AI drafted · Analyst reviewed
Drafted by AI from the full research note and reviewed by a Greyhound Research analyst before publication. The analysis and conclusions are Greyhound Research’s own.
At Microsoft Ignite on November 17, 2026, Microsoft gave AI agents infrastructure of their own on Azure, its cloud. Each agent can now get a sealed-off machine for every task, network rules that refuse by default, a saved copy of work the model has already done, and help choosing the right chips. A new route also lets companies run Microsoft’s Foundry models and agents on their own Kubernetes clusters, the systems most companies use to run software. Almost all of it is in preview, a test release without full service terms. Two pieces are generally available, meaning fully released: Cobalt 200 virtual machines now, and Container Apps Sandboxes since September.
Greyhound Research’s position is simple. Ignite made the machinery for running agents on Azure ready to pilot. What matters next is what each agent may touch, and under whose name. Microsoft’s security leaders call agents insiders, and they are right. A machine of its own stops an agent’s code from reaching the computer underneath. It does not stop an agent misusing access it already holds. Greyhound Research reads this as a familiar turn in a platform shift. The machinery is becoming a commodity faster than the rules for using it are being written, so the buyers likely to fare best will adopt what is released and write down each agent’s authority themselves, before Microsoft’s full terms arrive.
Three findings carry the argument. First, Microsoft is level with AWS and Google on running agents, behind AWS on written rules for each tool an agent calls, and ahead on everything around the agent, where its identity, security, and Microsoft 365 products already sit. Second, Context Cache, Microsoft’s feature for reusing prompt text, saves money only where agents reread the same long text. Because it keeps that text for days or weeks, it is also a new store of data to govern. Third, running Foundry on a customer’s own clusters matters most to regulated buyers, as long as each way of running it comes with a written split of duties.
The note places lock-in in an agent’s identity, memory, and records, not in the AI model it calls. It sets out a ten-move Greyhound CXO Runbook, with eleven exhibits, six tables, and ten questions for Microsoft. None of the moves waits on Microsoft. The first three are simple to state. Pilot on what is fully released. Register every agent as a privileged account before it gets any tool. Until Agent Grid is documented, hold limits on each tool in a gateway the company controls.
The evidence has limits. Most of what Microsoft announced is in preview. Searches of Microsoft Learn, its documentation site, found no pages for Agent Grid or the sandbox extension for its Kubernetes service through 00:45 UTC on November 19, 2026. Start-up speeds from Microsoft, AWS, and Google are each company’s own claims, measured in different ways. Greyhound Research names the signals that would change its view. Until they arrive, the advice holds: pilot now on what is fully released, and commit production only when documentation, service levels, and full-release terms are published.
In This Research NoteSection guide · AI drafted · Analyst reviewed
Twenty-three sections in nine parts, each with a short summary. Select a title to go straight to that section. Each section opens with a Greyhound Spotlight, the one fact in it a reader would not expect, and eighteen end with a block named The Deeper Read, which holds the fuller argument and the passages it rests on. Select any line in that block to open its full text.
- Part 1What Ignite Changed
- 1.1What Ignite Changed: Agents Get Machines Of Their Own
Microsoft gave AI agents their own machines, network rules, memory, and chip placement at Ignite. The five things an agent at scale needs, what Ignite added to each, and the test this note applies.
- 1.2What Microsoft Announced: Six Previews And One General Availability
What is in preview and what is fully released. Six infrastructure products are in preview, Cobalt 200 is fully released, and Container Apps Sandboxes has been since September. Microsoft’s own text puts the new sandboxes on its Kubernetes service.
- Part 2The Runtime And Its Limits
- 2.1Sandboxes At Scale: Speed Is Becoming Table Stakes
Microsoft, AWS, and Google all start sandboxes fast, measured in different ways. Speed is now the entry ticket. Proof linking each sandbox to an identity and a result is the open contest.
- 2.2Agents As Insiders: Isolation Is The Floor, Not The Ceiling
Microsoft is right to call agents insiders. Harm arrives by three routes, and a sandbox stops only one. Recent security findings point to the other two.
- 2.3Agent Grid And Identity: Where Authority Is Enforced
Agent Grid puts identity-based, refuse-by-default rules on how agents connect. How fine those rules go, and how each session is tied to an agent identity, is not yet documented.
- Part 3Memory And Placement
- 3.1Context Cache Economics: A Discount On Repetition
Caching discounts only the repeated part of what an agent sends. Other providers’ prices set the ceiling, Microsoft has published none, and keeping the cache for days is what is new.
- 3.2Context Cache Governance: A New Store With A Retention Date
Saved prompt text stays in the customer’s own Azure account until it has gone unused for up to 30 days, in two preview regions. It belongs in the data inventory before the cost plan.
- 3.3Confidential Computing: The Seal And The Reader
Ignite’s summary put three confidential computing offerings in preview, and Microsoft’s detailed post puts only the clean room there. All keep the cloud operator out, not the agent inside, and the proof is checked locally for now.
- 3.4AKS AI Manager: Placement Judgment For Scarce GPUs
AI Manager helps decide which chips each model should run on. Whether it is built on open-source projects Microsoft has not named decides whether a team can move away later.
- Part 4The Customer’s Own Ground
- 4.1Foundry On Customer AKS: Microsoft’s Catalog On The Customer’s Ground
The announcement that matters most to regulated buyers. Running duties and chip shortages pass to whoever runs the cluster.
- 4.2Shared Responsibility: The Seam Between Two Operators
Microsoft promises the same rules whether it runs the cluster or the customer does. Who patches, who answers for outages, and who keeps the logs are not yet assigned in writing.
- 4.3Capacity And Quota: The Constraint Behind The Platform Choice
Residency rules fix the region, and the region fixes which chips are available. Quota belongs at the start of planning.
- Part 5The Platform Contest
- 5.1Three Clouds Compared: Level On Runtime, Ahead On The Estate
Layer by layer, Microsoft is level with AWS and Google on running agents and on identity, behind AWS on written tool rules, and ahead on the wider estate.
- 5.2Lock-In And Exit: The Agent Carries What The Model Does Not
AI models are easy to swap. An agent’s identity, memory, and records are not, on any cloud. Open standards help at the edges.
- 5.3Buy Or Build: Buy The Commodity, Own The Decisions
Buy the machinery, the sandbox, and the model catalog. Own the permissions, approvals, and testing an auditor will ask about.
- Part 6The Risks And The Counter-Argument
- 6.1Risks To Watch: Where Agent Infrastructure Can Go Wrong
Six risks, from previews changing to disputes between operators, each with an early warning, an owner, and the Runbook move that answers it.
- 6.2The Strongest Objection: The First Wave Only Needs A Sandbox
The objection holds for running code and falls short on timing for everything else agents now do.
- Part 7What To Do, By Role
- 7.1The Greyhound CXO Runbook: Ten Moves, None Waiting For Microsoft
Ten moves for platform engineering, security, finance, data protection, regulated CIOs, the CIO, and procurement, each naming the assumption it corrects and who decides.
- 7.2Questions For Microsoft: What Buyers Will Ask Next
Ten questions only Microsoft can answer, on where sandboxes run, service terms, rule depth, identity, cache price and contents, open source, duties, capacity, and exit. Each is tied to a Runbook move.
- Part 8The Final Greyhound Standpoint
- 8.1Azure Agent Infrastructure: The Final Greyhound Standpoint
Running agents is ready to pilot, the wider estate is Microsoft’s advantage, and control over what agents may do is still to be written. The limits of the evidence, and the signals that would change the view.
- Part 9Appendix, Sources, And Disclosures
- 9.1Appendix: The Evidence Register, Source By Source
Every source behind this research note, with when it was read, what was taken from it word for word, and where that material is used, in one drop-down for each class of source, with the searches that found nothing.
- 9.2Sources And References: Every Document Behind Every Claim
Sixty-two sources in seven classes, from Microsoft’s announcements and documentation to AWS, Google, model providers, open-source projects, security research, and the research this note builds on, each also linked in the sentence that carries it.
- 9.3Disclosures: Independence, Financial, Intended Use, Methodology, Authoring, And Copyright
Independence first, then the Financial Disclosure, intended use, research method and its limits, authoring, and copyright terms.
Where To Start, By RoleReader guide · AI drafted · Analyst reviewed
Find your role and read its sections in the order shown. Each line takes you straight to the section.
Chief information officer
Chief information security officer
Head of platform engineering
Enterprise architect
FinOps lead and chief financial officer
Data protection officer
Head of procurement
Part 1What Ignite Changed
What Ignite Changed: Agents Get Machines Of Their Own
Greyhound SpotlightMicrosoft filed four of its new agent products under one heading for running AI agents in production. All four are in preview, a test release without full service terms.
Ignite 2026 was the week Microsoft gave AI agents infrastructure of their own on Azure. Microsoft announced four new pieces under one heading and a senior executive described two more. Together they give each agent a sealed-off machine, network rules, saved memory, and help finding the right chips. For a CIO, the change is that agents now have to be run and governed like any other workload.
What is new is the unit of planning. Until now, most enterprises bought AI as a feature inside an application and governed it through that application’s settings. Ignite treats each agent as a workload with its own machine, network rules, memory, and chips, which places agents inside the platform team’s estate beside databases and web services. Greyhound Research expects agent numbers to grow faster than the review processes built for ordinary workloads, much as virtual machine sprawl outran change control a generation ago. The second-order effect falls on ownership. A workload needs an owner on the operations side as well as a sponsor in the business, and many agent pilots today have only the second.
Exhibit 1 lists the five things an agent needs to work at scale. Only the first, a fresh machine for each session, is generally available today. Generally available means fully released, with support and service terms.
Greyhound Research · Exhibit 1 · The chain of agent scale
Only the first of five links is generally available
The five links an agent at scale depends on, read against Microsoft’s written Ignite record and the documentation beside it, November 17 and 18, 2026.
Greyhound Research’s judgment: buy the first link now, and treat the four that decide authority as go-live conditions, not features.
Exhibit 1 sets out, as five numbered steps, the links an agent at scale depends on, and shows that only the first is generally available. Step one, start, a disposable machine per session, is generally available through Container Apps Sandboxes, with the Agent Sandbox extension for AKS in preview. Step two, identity, is partly written: each Foundry hosted agent gets its own Entra ID, while the binding of an AKS sandbox session to an Entra Agent ID is not documented. Step three, tool reach, is described in preview through Agent Grid, without stating whether policy reaches the individual tool call. Step four, working memory, is described in preview as durable cached context in the customer’s tenant, deleted after 1 to 30 days unused, without a price or a path for erasure on request. Step five, one record joining prompt, tool call, and changed record, is not yet written in any document read. The links and grades are Greyhound Research analysis, and the reading rests on documents published by November 19, 2026.
Only the first link is a product a company can buy now. The other four are partly written, in preview, or not written at all. Those four decide what an agent may do. The exhibit cannot show how they will look at full release. Until then, each one belongs in a pilot’s go-live conditions, not in a production plan.
Greyhound Standpoint:
At Greyhound Research, we believe Ignite changed the category agents belong to more than it changed any single product. Microsoft gave agents what every production workload eventually receives: computing of their own, network rules, a place to keep memory, and help finding scarce chips. Once software has a machine of its own, the machine stops being the hard question. What the software may read, change, and spend becomes the question, because that is where the money and the liability sit.
Practitioners have watched this sequence before. When server virtualization arrived, the hypervisor became a commodity within a few years, and the lasting value and the lasting risk moved to the management, identity, and licensing layers above it. Agent infrastructure is moving along the same path at a faster pace. Microsoft deserves credit for assembling the lower layers quickly, and for proving one of them at its own scale.
The evidence has a clear limit. Only one of the five links in Exhibit 1 is generally available, and most of the rest are previews whose terms can still change. The view would weaken if Agent Grid arrived with rules fine enough to govern each tool call, because authority would then ship with the machine. Until then, CIOs should list each preview link as a go-live condition in pilot plans, and keep it out of production budgets at the next architecture review.
The Deeper Read · About 2 minutes
Select a question to open the full answer
1What exactly did Microsoft announce, and in whose words?
In Microsoft’s words. Greyhound Research reads Microsoft Ignite 2026 as the week Azure gave AI agents infrastructure of their own. Microsoft’s Book of News for Ignite 2026 places four new items under one heading, the Azure infrastructure foundation for production AI applications and agents: an Agent Sandbox extension for AKS, Agent Grid, a path for Foundry models and agents onto customers’ own AKS clusters, and flex nodes for AKS. Jeremy Winter, Corporate Vice President and Chief Product Officer for the Azure Platform, describes two more in his Ignite post, reuse of processed prompt context and AKS AI Manager.
The start is the one link that is already a product: Container Apps Sandboxes is generally available, and the AKS extension is in preview. Each link after it is partly written, in preview, or not yet written, and those are the links that decide what an agent may do. What the exhibit cannot show is how the later links will read at general availability. Until then, every step past the first is a condition to write into go-live plans.
2What changes when agents become workloads?
From feature to workload. The shift is from agents as features inside applications to agents as workloads with needs of their own. An agent that writes and runs code needs somewhere disposable to run it. Calling tools requires a policy on what it may reach. Long-running work needs context that survives a pause, and the models behind it need GPUs chosen for them. Ignite gave each of those needs a named Azure product, most of them in a form an enterprise can try this quarter.
3Why does Microsoft’s insider framing shape the whole research note?
An insider is defined by access. The same week, Microsoft’s security leadership pressed a second idea. Aleš Holeček, President and Chief Architect of Microsoft Security, wrote on November 17 that “Agents are the new insiders.” That framing shapes every section that follows, because an insider is defined by the access it holds, not by the room it works in. A machine of its own settles where an agent’s code runs. Authority, meaning what the agent may read, change, and spend, settles what it can do.
What the insider label tells a buyer. Holeček’s quoted line ties Entra Agent ID and Conditional Access to the agents Microsoft discovers, which is the language of identity governance, not of infrastructure. That tells a buyer which team Microsoft expects to govern agents: the one that already reviews human access. It also sets a test the buyer can apply. If an agent can be discovered and given an identity, it can be reviewed and have rights removed, and a buyer should ask to see that last step working.
In the source’s own words
- Microsoft Security Blog, Holeček at IgniteDiscovered agents receive a Microsoft Entra Agent ID and can be governed with Conditional Access.Source
4How does this research note connect to the Microsoft 365 outage note?
The same test, applied to agents. Ignite also opened days after the Microsoft 365 outage that Greyhound Research examined on November 15 in Microsoft 365 Outage – When The Suite Goes Dark, Restoration Order Becomes The Control. That note held that consolidation on one suite remains the right strategy for most enterprises while the standard terms around it fall short. This research note applies the same test to agent infrastructure. Integration is worth buying when the terms, the evidence, and the way out are written.
Part 1What Ignite Changed
What Microsoft Announced: Six Previews And One General Availability
Greyhound SpotlightMicrosoft’s Ignite summary puts the new agent sandboxes on its Kubernetes service. Its Kubernetes team’s Ignite post adds that their serverless option runs on Azure Container Instances, its other container product.
What a buyer can sign depends on whether a product is in preview or fully released. Microsoft’s Ignite summary is plain about it. Everything is in preview except Cobalt 200, a new family of Azure virtual machines. That covers the sandbox extension, Agent Grid, the Foundry route to customer clusters, and flex nodes. Context caching and AI Manager are previews too, according to their own repository and launch post.
The status split matters more than usual because the two released products sit at the base of the stack, while everything that would govern an agent sits above them in preview. In earlier cloud cycles, a buyer could often adopt the compute first and wait for the management layer at little risk. Agents change that, because the governing pieces decide what the released machinery is allowed to do. Greyhound Research anticipates several previews moving toward full release in the coming months, though no dates are published. The consequence that tends to be overlooked sits in budgeting. A preview with no price cannot be costed, so a business case that includes one carries an estimate, not a number.
Table 1 lists each product with its status. Six are in preview with no published price, one is announced as generally available, and one has been generally available since September.
Greyhound Research · Table 1 · Status ledger
Six items in preview with no published price
Microsoft’s Ignite 2026 Azure infrastructure items by status, with what is written and what is not, November 19, 2026.
Sign only the two released rows. Treat every preview row as a plan until its status, price, and service level are in writing.
| Item | Status | What Microsoft has written | Not yet written |
|---|---|---|---|
| Agent Sandbox extension for AKS | Preview | “run agent-written code in sandboxes isolated at the virtual-machine level”Pre-warmed pools and a serverless option, per Winter | Learn documentation, regions, price, service level |
| Agent Grid | Preview | “centralized policy enforcement”Identity-driven, default-deny connections, per Winter | Depth of policy, documentation, price |
| Context caching, Azure Context Cache | Preview | Durable cached context in the customer’s tenantCentral US and Sweden Central, 1 to 30 days | Rate, write premium, deletion path |
| AKS AI Manager | Preview | Model catalog, GPU placement, cost by namespaceAPI version 2026-09-02-preview | Open-source base, price |
| Foundry to customer AKS | Limited preview | Open-weight models and agents onto the customer’s own AKSGPUs and regions chosen by cost, latency, and residency | Duties per mode, service level, date |
| Flex nodes for AKS | Preview | “extend the same AKS clusters to customer servers outside Azure” | Supported configurations for Foundry |
| Azure Cobalt 200 virtual machines | Generally available | 14 regions nowUp to 50% better than Cobalt 100, and up to 1.5 times on agent tool execution | Year-end regions: 23 in the Azure post, 24 in the Book of News |
| Azure Container Apps Sandboxes | Generally available | Prewarmed pools, thousands concurrentNo CPU or memory fees while stopped | Rates not on the overview page |
| Hosted agents in Foundry Agent Service | Status not on page | Per-session VM-isolated sandboxesAn Entra ID per agent, 31 listed regions | Status label on the documentation page |
| ACI Sandboxes | Technology preview | “Emerging Technology Preview”Demonstrated at Build 2026. The AKS extension’s serverless option runs on ACI | Status as a product of its own |
| Maia 300 | Announced | “40 percent more tokens per watt”Against leading GPU-based systems, per Nadella | Availability and regions |
Table 1 lists Microsoft’s Ignite 2026 Azure infrastructure items by status, and shows six items in preview with no published price and one item announced as generally available. In preview: the Agent Sandbox extension for AKS, Agent Grid, context caching, AKS AI Manager, Foundry to customer AKS, and flex nodes, each with what Microsoft has written and what is not yet written, such as documentation, price, depth of policy, deletion path, duties per mode, and service level. Generally available: Azure Cobalt 200 virtual machines, in 14 regions with 23 or 24 by year-end depending on the document, and Azure Container Apps Sandboxes. Context: hosted agents in Foundry Agent Service, ACI Sandboxes as an emerging technology preview, and Maia 300 as announced. The statuses are as Microsoft published them through November 19, 2026.
Only two rows can be signed today: Cobalt 200 and Container Apps Sandboxes. Everything above them can still change before launch. None of the six preview products has a published price. The table grades what Microsoft has written about status, not how well the products work. Budgets belong on those two rows for now.
Greyhound Standpoint:
In Greyhound Research’s reading, the word that should carry most weight in Microsoft’s Ignite news is preview, and the calendar deserves as close a reading as the feature list. A preview is a test drive. Its features can change shape before launch, a service-level agreement usually does not apply, and price, region list, and support terms are often unpublished. All three conditions hold today for Agent Grid, the sandbox extension, AI Manager, and the Foundry route to customer clusters.
The pattern is familiar from every large cloud transition. Services announced on a conference stage have often been renamed, repriced, or reshaped before general release, and buyers who signed against the keynote then renegotiated against the product. That history argues for discipline, not for distance. Microsoft has earned some credit on timing. Sandboxes shown at Build 2026 became a generally available Container Apps product within months, and early testers shape what ships. The order of arrival matters as well. Agent Grid’s release date will matter more than any start-up speed, because the new sandboxes are governed through it. Published full-release terms for Agent Grid and the sandbox extension would shorten the case for waiting.
Engineering teams should test now on non-production data, against written go-live conditions. Procurement should hold any budget commitment beyond 2027 until full-release terms are published, and revisit that position at each renewal checkpoint.
The Deeper Read · About 4 minutes
Select a question to open the full answer
1What exactly does Microsoft’s text say about status?
The record in detail. Status decides what a buyer can sign, so Greyhound Research counts what Microsoft has written about it. The Book of News is plain on the point: “All capabilities are in preview except Azure Cobalt 200, which is generally available.” That sentence covers the Agent Sandbox extension for AKS, Agent Grid, the Foundry path to customer AKS, and flex nodes. Winter’s post gives no separate status for context caching or AI Manager, while the Azure Context Cache repository and the Kubernetes Center post of November 16 both describe previews.
The two generally available rows, Cobalt 200 and Container Apps Sandboxes, are the ones a buyer can sign today. Everything above them can change shape before launch, and none of the six preview rows carries a published price. The table grades what Microsoft has written about status, not how well the products work. Commitments belong on those two rows until Microsoft publishes terms for the others.
2What can buyers use today, beside the announcements?
Two products already in production. Azure Container Apps Sandboxes is generally available, and Mike Hulme, General Manager of Azure Product Marketing, wrote on September 23 that more than a million sandboxes a day run in production across Microsoft. Foundry’s hosted agents already “run in per-session VM-isolated sandboxes,” each with its own Entra identity, in 31 listed regions.
A third, still a demonstration. ACI Sandboxes is described as an “Emerging Technology Preview” demonstrated at Build.
In the source’s own words
- Microsoft’s Book of News, Azure infrastructure foundationAll capabilities are in preview except Azure Cobalt 200, which is generally available.Source
- Microsoft Community Hub, ACI Sandboxes pageEmerging Technology PreviewSource
- Microsoft Learn, hosted agents in FoundryHosted agents run in per-session VM-isolated sandboxes.Source
3What does the written record say, and not say, about the new sandboxes?
Where they run. Two clarifications follow from the written record. Microsoft’s Ignite text places the new sandboxes on AKS, and neither the Book of News nor Winter’s post names Azure Container Instances. Microsoft’s AKS Ignite post, read on November 19, fills that gap. It says the extension is “built on the open-source Kubernetes Agent Sandbox project” and offers “Kata microVM isolation on your cluster's own nodes” or “a serverless runtime enabled by Azure Container Instances (ACI).” Container Instances therefore sits beneath the extension’s serverless option, not beside it as a separate Ignite product.
Two claims, not one. Its published figures are separate claims as well. The Container Apps documentation says sandboxes are “provisioned from prewarmed pools for near-instant availability” and that the service “bursts to thousands of concurrent sandboxes on demand.” Of the two, one is a per-sandbox start time and the other a ceiling on concurrency. No Microsoft document read for this research note joins them into one figure. A search for Microsoft statements on ACI sandboxes at Ignite 2026, with start-up and scale figures, run on November 18, 2026, up to 02:06 UTC, found no Microsoft clarification, only the ACI preview post from the Build period, the Container Apps documentation, and secondary coverage. The AKS post, read on November 19, answers where the serverless option runs and gives no start-up or scale figure.
What the two clarifications tell a buyer. The record is precise once read in full, and precision changes planning. The serverless option runs on Container Instances beneath the AKS extension, so a team choosing between Microsoft’s sandbox routes is choosing between operating models, not between unrelated products. The warm-pool and burst figures are separate claims, so neither should be quoted as a measure of the other. A buyer comparing offers should ask which route each figure describes before it enters a business case.
In the source’s own words
- Microsoft’s Book of News, Azure infrastructure foundationAgent Sandbox extension for AKS will run agent-written code in sandboxes isolated at the virtual-machine level.Source
- Microsoft Community Hub, AKS at Ignite 2026a serverless runtime enabled by Azure Container Instances (ACI)Source
- Microsoft Learn, Container Apps Sandboxes overviewSandboxes are provisioned from prewarmed pools for near-instant availability.Source
- Microsoft Learn, Container Apps Sandboxes overviewThe service bursts to thousands of concurrent sandboxes on demand.Source
4What sits underneath, in silicon?
Cobalt 200 and Maia 300. Rani Borkar and Arun Kishan wrote on November 17 that Cobalt 200 virtual machines are generally available in 14 Azure regions, with “up to 50% better performance than Cobalt 100” and gains of up to 1.5 times for AI agent tool execution. Their post puts availability at 23 regions by the end of 2026, while the Book of News says 24. Satya Nadella’s Ignite post says Maia 300 “will deliver 40 percent more tokens per watt compared with leading GPU-based systems.” Both are Microsoft’s own claims. The Cobalt post also quotes a customer, Luc Choubert, VP Engineering at Amadeus: “In our benchmark lab, Azure Cobalt 200 improved flight search throughput by more than 30% compared with Cobalt 100,” a benchmark result, not a production one.
In the source’s own words
- Microsoft Azure Blog, Cobalt 200 announcementgenerally available today in 14 Azure regionsSource
- Microsoft Azure Blog, Cobalt 200 announcementup to 50% better performance than Cobalt 100Source
- Microsoft Azure Blog, Cobalt 200 announcementup to 1.5 times for AI agent tool execution workloadsSource
- Microsoft Azure Blog, Cobalt 200 announcementAvailability will expand to 23 regions by the end of 2026.Source
- Microsoft’s Book of News, Azure Cobalt 200deployed in 24 Azure regions by the end of the yearSource
- Nadella’s Ignite post, on Maia 300We just announced Maia 300, which will deliver 40 percent more tokens per watt compared with leading GPU-based systems.Source
Part 2The Runtime And Its Limits
Sandboxes At Scale: Speed Is Becoming Table Stakes
Greyhound SpotlightMicrosoft, AWS, and Google each publish a start-up speed for their sandboxes. No two measure the same thing.
A sandbox is a sealed-off machine where an agent can run code it wrote without reaching anything else. Microsoft describes its new sandboxes on Azure Kubernetes Service (AKS) as getting a virtual machine of their own for every session. Machines are kept ready in advance for a fast start. That protects the computer underneath from untrusted code, the worst failure when agents write and run software. Agents that write software, analyze data, or automate internal work face that risk most often today.
For buyers, the practical news is that a fast, sealed-off start is no longer a reason to choose one cloud over another. Each of the three large providers now publishes a start-up claim, and each measures it differently, so the figures cannot settle a purchase. Infrastructure features usually mature this way: what sets a vendor apart one year becomes a line in every request for proposal the next. Greyhound Research expects the comparison to move toward evidence over the coming year, meaning which identity each session ran under and what it changed. The effect most coverage misses is on procurement scoring. Evaluation templates that weight start-up speed heavily will reward a feature that has already stopped separating the field.
Exhibit 2 sets each published start-up claim against the stage of a sandbox’s life it measures. A warm pool is a set of machines kept running in advance. Burst means how many can run at once. Five claims turn out to measure four different things.
Greyhound Research · Exhibit 2 · What each start-up claim measures
Five start-up claims, four measurements, no common clock
Microsoft’s, AWS’s, and Google’s published start-up and scale claims, placed against the stage of a sandbox’s life each one measures.
Speed no longer separates the clouds. Choose on the identity and evidence behind each session, not on the stopwatch.
Exhibit 2 is a matrix of four stages of a sandbox’s life against five published start-up claims, and shows that no two claims measure the same thing. The stages are pool kept warm, one sandbox handed out, session ready with image loaded, and many running at once. Container Apps Sandboxes claims near-instant availability from prewarmed pools, covering the first two stages, and bursts to thousands of concurrent sandboxes, the fourth stage. The AKS extension describes pre-warmed pools for fast startup in preview, with no figure. AWS says AgentCore Runtime sessions start in under 2 seconds, covering allocation and session readiness. Google says GKE Agent Sandbox can allocate 300 sandboxes per second per cluster, with 90% of allocations within 200 milliseconds, covering the first two stages. A row beneath counts, for each claim, the stages it measures and the stages it describes without a figure. All are company claims on different bases and are not directly comparable, and the stage mapping is Greyhound Research analysis.
These figures are not a league table. Each company times something different and publishes its own figure. Every major cloud now offers a fast, sealed-off start as standard. Speed has become the price of entry rather than the thing that sets a provider apart. A buyer comparing platforms learns more by asking which identity each session runs under.
Greyhound Standpoint:
Greyhound Research’s assessment is that Microsoft has made the visible part of agent scale into a product and left the harder part with its customers, as AWS and Google have also done. Once every serious cloud offers a fast, sealed-off start, speed stops deciding purchases and the contest moves to what is still scarce. What is scarce here is proof: a record tying each sandbox to the agent behind it, the identity it acted under, and what it changed in the business. No cloud has finished that work.
Experienced buyers will recognize the curve. Start-up time, like virtual machine boot time a generation earlier, is a benchmark vendors compete on until everyone clears it, after which the operating record matters more than the stopwatch. Microsoft deserves credit for running more than a million sandboxes a day, a volume few companies will reach, which makes the Container Apps product a sound base for pilots now.
The comparison has a limit. Each company times something different, so Exhibit 2 supports no ranking. The Kubernetes route will be ready for production when Microsoft documents which agent identity each session runs under, commits to a service level, and dates Agent Grid’s full release. AWS’s re:Invent, opening on November 30, could reset the comparison. Platform engineering leads should pilot this quarter, and keep the preview out of production until it carries documentation and service terms.
The Deeper Read · About 3 minutes
Select a question to open the full answer
1How exactly does Microsoft describe the sandbox, and what does a quick start protect?
The design. A fast start buys a platform team one thing, and Greyhound Research tests the sandbox news against it. Winter describes the AKS extension as running agent-generated code with “virtual machine (VM)-level isolation for every session, pre-warmed pools for fast startup,” and a serverless option. Giving every session a throwaway machine means untrusted code cannot reach the host, which is the gravest failure in code execution. Agents that write software, analyze data, or automate internal tasks face that hazard most often today.
These figures are not a league table. A warm pool, an allocation rate, a full session, and a concurrency ceiling are four different clocks, and each company publishes its own. What the exhibit does show is that every major cloud now treats a fast, isolated start as standard, so start-up speed is becoming the entry condition rather than the point of difference. Buyers comparing platforms gain more by asking which agent identity each session runs under.
2What do Microsoft’s rivals ship, and how do their figures compare?
The same pattern elsewhere. Microsoft’s rivals already ship the same pattern. AWS said on November 12 that sessions on its next-generation AgentCore Runtime “start in under 2 seconds,” each in “its own hardware-isolated Firecracker micro virtual machine.” Google said on May 20 that GKE Agent Sandbox was generally available and could “allocate 300 sandboxes per second, per cluster, at sub second latency, with 90% of allocations completing in 200 milliseconds.” Cloudflare offers sandboxes for agent code as well, so AWS, Google Cloud, Azure, and Cloudflare now all ship one, and no two are built the same way.
Different clocks. Each vendor measures something different: a single start from a warm pool, an allocation rate per cluster, a full session start that includes the container image, or how many sandboxes can run at once. The comparison that survives is narrower and more useful, which is that every major cloud now treats a fast, isolated start as standard.
What the rivals’ own words tell a buyer. AWS’s quoted figure counts a full session start, and Google’s counts allocations per cluster each second. Each describes something real, and neither can be set beside Microsoft’s warm-pool start without a test on the buyer’s own workload. The quotes also show that every vendor now sells an isolated machine per session as standard. On isolation the field has converged, so the questions that still separate the offers concern identity, audit, and the cost of each session.
In the source’s own words
- AWS Machine Learning Blog, November 12every session runs in its own hardware-isolated Firecracker micro virtual machine (MicroVM)Source
- Google Cloud Blog, May 20GKE Agent Sandbox is now generally availableSource
- AWS Machine Learning Blog, November 12Sessions start in under 2 seconds, independent of image size and concurrency.Source
- Google Cloud Blog, May 20allocate 300 sandboxes per second, per cluster, at sub second latency, with 90% of allocations completing in 200 milliseconds.Source
3What does a quick start leave unsettled?
Everything after the start. A quick start settles nothing that happens next. Each running agent still needs a name the security team can trace and revoke, access to no more tools than its task requires, memory that outlasts an interruption without turning into an unmanaged store, and a single trail from prompt to tool call to altered record. That trail is the weakest part of every cloud’s offer.
What CIOs prioritize. Greyhound Research reported from the Greyhound CIO Pulse 2025 that 61% of CIOs globally prioritize AI observability over raw model performance when evaluating GenAI investments.
In the source’s own words
- Greyhound Research, Microsoft Build 2025 noteAccording to the Greyhound CIO Pulse 2025, 61% of CIOs globally now prioritise AI observability over raw model performance when evaluating GenAI investments.Source
4What is the strongest case for Microsoft here?
Proven under load. Microsoft’s defenders have a strong point here. More than a million sandboxes a day across Microsoft means the Container Apps design has been exercised under real production load, and that is the best argument for adopting the generally available product now.
The AKS extension is a different case. As a preview, it carries no service-level commitment yet, and searches of Microsoft Learn through November 19, 2026, at 00:45 UTC found no documentation, region list, or price for it.
In the source’s own words
- Winter’s Ignite post, on the Agent Sandbox extensionWith the Agent Sandbox extension for AKS, agent-generated code runs in secure, managed sandboxesSource
- Winter’s Ignite post, on the Agent Sandbox extensionvirtual machine (VM)-level isolation for every session, pre-warmed pools for fast startup,Source
- Winter’s Ignite post, on the Agent Sandbox extensiona serverless option with no infrastructure to manage.Source
Part 2The Runtime And Its Limits
Agents As Insiders: Isolation Is The Floor, Not The Ceiling
Greyhound SpotlightEchoLeak, a flaw in Microsoft 365 Copilot that Microsoft rated 9.3 on CVSS, the standard 10-point severity scale, let data leak without anyone clicking anything. No sandbox escape was involved.
Microsoft’s security leaders call AI agents insiders, and Greyhound Research agrees. The label has a consequence the sandbox news does not reach. Harm from an agent comes by three routes. Its code can break out of its machine. It can be tricked, by text it reads, into misusing access it already has. Or it can be given more access than its task needs from the start. A sandbox deals with the first route. Network rules that refuse by default narrow the second. Only identity and access controls fix the third.
Calling agents insiders moves the problem from the security perimeter to the access review, and that is a different team with a different calendar. Platform and network engineers own the perimeter and workload controls. Identity teams own entitlements and, in most companies, review them on a cycle built for people who change roles a few times a year. Agents can be created, copied, and granted access in minutes. On Greyhound Research’s reading, many organizations will find their first agent risks in the access review rather than the security console. The follow-on effect is on accountability for rights. Every agent needs a person who answers for what it may reach, and that person is usually not the engineer who built it.
Exhibit 3 sets the three routes against four controls that Microsoft and its rivals document. Default-deny egress means outbound traffic is blocked unless allowed. A separate machine per session reaches only the first route.
Greyhound Research · Exhibit 3 · Routes of harm against controls
The sandbox reaches only the first of three routes
Greyhound Research’s reading of how agent harm arrives, set against four controls Microsoft and its rivals now document.
A sandbox is the floor. Narrow rights, approval, and identity are what stop an insider, and they belong in the design now.
| Route of harm | Virtual machine per session | Default-deny egress, credentials held outside | Identity per agent, narrow tool rights | Human approval on irreversible actions |
|---|---|---|---|---|
| 1. Code escapes its hostAgent-written code breaks out onto the machine underneath | Answers | Narrows | Does not reach | Does not reach |
| 2. Steered into misusing held accessText the agent reads turns access it legitimately holds against the enterprise. EchoLeak, CVE-2025-32711, involved no sandbox escape. | Does not reach | Narrows | Narrows | Answers |
| 3. More access than the task needsThe agent is granted rights beyond its task from the start | Does not reach | Does not reach | Answers | Narrows |
Exhibit 3 is a matrix of three routes of agent harm against four controls, with a heavy rule after the first control marking where a sandbox’s protection stops, and shows that a virtual machine per session reaches only the first route. Route one, code escaping its host, is answered by a virtual machine per session and narrowed by default-deny egress. Route two, an agent steered into misusing access it holds, as in EchoLeak, is not reached by the virtual machine, is narrowed by egress rules and by narrow tool rights, and is answered for irreversible actions by human approval. Route three, more access than the task needs, is answered by an identity per agent with narrow tool rights and narrowed by approval. The grades are Greyhound Research analysis, not measured data.
A stronger wall around the machine would change none of the public findings. EchoLeak needed no escape. An agent that can write to a ledger can do harm from inside its own sandbox, simply by obeying a poisoned instruction it has read. The grades are Greyhound Research’s reading, not measured data. They show that only narrow rights, human approval, and identity stop the second and third routes.
Greyhound Standpoint:
Greyhound Research holds that Microsoft has named the risk accurately and built the right first barrier, and that the barrier cannot carry the whole risk. A sandbox is a locked room. An insider becomes dangerous through the keys it carries out of that room: the records it may read, the systems it may change, and the credentials it holds.
Security leaders have governed human insiders for decades. Privileged access programs worked where they reviewed entitlements, kept privileges narrow, and removed rights nobody used. Programs that leaned on monitoring alone tended to record incidents rather than prevent them. Service accounts offer the cautionary precedent, because rights granted for convenience during a project usually outlived it. Microsoft deserves credit for the fuller answer it is building, with an Entra identity for each hosted agent and Dynamic Agent Risk above it. The limit is that EchoLeak and the National Cyber Security Centre’s warning describe a class of risk, not how often it occurs in production. The view would soften if regular entitlement reviews for agents became a documented, built-in Microsoft control.
CISOs should treat every agent as a privileged account from its first day, with an accountable person, a written list of permitted tools, approval before irreversible actions, and a quarterly review that removes unused rights. That work belongs in this quarter, before full release, because rights granted during a pilot tend to survive into production.
The Deeper Read · About 3 minutes
Select a question to open the full answer
1How do the three routes work, and what does each control reach?
The three routes in detail. Greyhound Research reads Microsoft’s insider framing as right, and as carrying a consequence the sandbox news does not reach on its own. Harm from an agent arrives by three routes. Code can escape its host. An agent can be steered, through text it reads, into misusing access it legitimately holds. Or it can be granted more access than its task needs from the start. A virtual machine per session deals with the first route. Default-deny networking narrows the second, because exfiltrated data has fewer places to go. The third yields only to identity and entitlement.
None of the public findings would change with a stronger hypervisor. EchoLeak involved no sandbox escape, and an agent with write access to a ledger can do harm from inside its own sandbox, simply by acting on a poisoned instruction it has read. The grades are Greyhound Research’s reading, not measured data, and they do not rank the controls by cost. What they show is that the second and third routes yield only to narrow rights, approval, and identity. Greyhound Research made the same point in September 2026, after an agent containment failure elsewhere: “A sandbox can protect the host while failing to contain the work.”
2Where does Microsoft’s insider framing come from?
Three senior voices. Satya Nadella, Chairman and Chief Executive Officer of Microsoft, posted an essay in October titled “Models as Insider Risks in the Super Intelligence Era,” in which he wrote, “We need to separate the supply of intelligence from the authority over it.” Hayete Gallot, Executive Vice President of Microsoft Security, wrote on October 19 that internally deployed agents “are agentic insiders” and that they “are frequently overprivileged.” Holeček compressed the idea into five words at Ignite.
What three executives agreeing tells a buyer. When the chief executive and two security leaders use the same frame within weeks, it reads as a stated position rather than a passing phrase. Gallot’s word “overprivileged” is the most useful for buyers, because it names a condition they can measure in their own estates. Nadella’s line on separating intelligence from authority describes the design goal. A buyer can hold Microsoft to that goal by asking which of its products remove unused agent rights, not only which ones detect misuse.
In the source’s own words
- The Official Microsoft Blog, Gallot on October 19They are agentic insiders.Source
- The Official Microsoft Blog, Gallot on October 19These agents are frequently overprivilegedSource
- Nadella’s October essayWe need to separate the supply of intelligence from the authority over itSource
- Microsoft Security Blog, Holeček at IgniteAgents are the new insiders.Source
3What does the public evidence show?
The second and third routes. Public evidence points to the second and third routes. EchoLeak, CVE-2025-32711, which Microsoft rated 9.3 on CVSS in June 2025, let data be disclosed from Microsoft 365 Copilot without user interaction. Microsoft fixed it on its own side, and users had nothing to install. No sandbox escape was involved. The UK National Cyber Security Centre warned in December 2025 that these systems draw no inherent distinction between data and instruction.
In the source’s own words
- The UK National Cyber Security Centre’s blog, December 8, 2025Current large language models (LLMs) simply do not enforce a security boundary between instructions and data inside a prompt.Source
4Which of Microsoft’s controls reach the second and third routes?
Microsoft’s own design. Microsoft’s own design shows it understands the point. Its egress documentation for Container Apps Sandboxes recommends a starting posture of “default_action = Deny plus an explicit allow list,” and lets outbound calls be authenticated and routed “without the sandbox ever needing to hold the credentials directly.” Each hosted agent in Foundry gets “its own dedicated Microsoft Entra ID (agent identity).” Holeček writes that discovered agents receive an Entra Agent ID and can be governed with Conditional Access, and that Dynamic Agent Risk is available in preview. Those controls are aimed squarely at the third route.
What holds inside the sandbox. What holds there is tightly scoped tool rights, a person’s sign-off before anything irreversible, and an identity security operations can switch off without waiting on the application team. Per-sandbox default-deny networking remains essential, since it closes most of the exits data could take. It is the starting level of control, not the whole answer.
In the source’s own words
- Microsoft Security Blog, on Dynamic Agent RiskAvailable in preview.Source
- Microsoft Learn, Container Apps Sandboxes egressdefault_action = Deny plus an explicit allow listSource
- Microsoft Learn, Container Apps Sandboxes egressTransform and Rewrite together let you front a sandbox's outbound calls with policy that authenticates and routes them without the sandbox ever needing to hold the credentials directly.Source
- Microsoft Learn, hosted agents in Foundryits own dedicated Microsoft Entra ID (agent identity)Source
Part 2The Runtime And Its Limits
Agent Grid And Identity: Where Authority Is Enforced
Greyhound SpotlightThe new sandboxes are governed through Agent Grid. Searches of Microsoft’s documentation site found no pages for it through 00:45 UTC on November 19, 2026.
Agent Grid is the piece that governs the new sandboxes. Microsoft says it will apply central rules to how agents, models, tools, and Azure services talk to each other. Rules follow identity rather than network address, and the default answer is no. That instinct is sound.
The open question is how fine the rules go. Letting an agent connect to the HR system is a coarse rule. Allowing it only to read leave balances for one region, and not to write, is a fine one. Fine rules limit the damage when an agent is misled. Microsoft’s Ignite description covers connections and access to resources, not individual tool calls.
For a buyer, Agent Grid is the piece that would turn Azure’s agent machinery from a place to run code into a place to govern work. The pattern resembles the early years of central sign-on, when a single login arrived well before consistent permissions inside each application. A rule at the connection decides which systems an agent can reach, and a rule at the tool call decides what it may do once there. Greyhound Research expects Microsoft to document more of Agent Grid before full release, though no date is stated. Until then, design teams face a practical choice: wait for the platform rule, or build the fine rule themselves and accept that the two may later overlap.
Exhibit 4 fills in four written agent policy designs to the finest level each one documents. A tool request is a single action an agent asks a tool to take. Host, path, and method mean the web address and the type of request. AWS and Google now document rules at the tool-request level, and only the AWS rules are generally available.
Greyhound Research · Exhibit 4 · Depth of written policy
Only AWS has generally available rules for each tool an agent calls
Four agent policy designs, each filled to the finest level its own documentation states, November 19, 2026.
Agent Grid is described at the connection. Until a tool-level rule is documented, keep the fine rule in a gateway the company controls.
Exhibit 4 shows four agent policy designs as cards, each listing five levels of policy depth from finest to coarsest, arguments scope and time, individual tool request, host path and method, service or resource, and network destination, marked documented or not stated, and shows that AWS and Google document policy at the individual tool request, with only AWS generally available. Agent Grid, in preview, is documented to the service or resource level. Container Apps Sandboxes egress, generally available, reaches host, path, and HTTP method. AWS AgentCore Policy, generally available since March 3, 2026, reaches the individual tool request. Google’s Agent Gateway, whose page was updated on November 18, 2026, documents tool-level rules read from MCP requests without a single release label, and GKE Agent Sandbox, generally available, documents a default-deny network policy at the destination level. A level marked not stated may still be supported in documentation not yet published.
Microsoft’s own Container Apps already filters outbound traffic by web address and type of request, which is finer than Agent Grid’s Ignite description claims. A level marked not stated may still exist once Microsoft documents Agent Grid. The exhibit records what is written, not what the product can do. Until the documents exist, the fine rules have to live somewhere the company controls.
Greyhound Standpoint:
As Greyhound Research sees it, Agent Grid is aimed at the right layer but described too thinly for a buyer to rely on yet. Identity-based rules that refuse by default are a sound base, and Microsoft’s identity products give it more to build on than any rival has. Its value now rests on two things: whether a rule can reach an individual tool call, and whether each sandbox session carries a documented Entra Agent ID.
The gap between those levels is the gap between a building pass and a signing authority. Letting an agent reach the payroll system is a coarse rule. Letting it read one cost center without writing is a fine one, and fine rules limit the damage when an agent is misled. Network teams learned this when address-based firewalls gave way to application-aware controls. On the written record, rivals moved first. AWS made tool-level rules generally available in March, and Google now documents tool-level rules in Agent Gateway without a single release label. A missing document is not a missing capability, and a detailed Agent Grid page could change this reading.
Security architects should get three written answers before Agent Grid enters a production design: the finest thing a rule can name, how a session binds to an agent identity, and where credentials are added. Until then, tool-level limits belong in the application or in a gateway the company controls.
The Deeper Read · About 3 minutes
Select a question to open the full answer
1What exactly does Microsoft say Agent Grid does, and how fine do the rules go?
In Microsoft’s words. Greyhound Research reads Agent Grid as the component on which governance of the new sandboxes depends. The Book of News says it “will secure interactions among agents, models, tools and Azure services with centralized policy enforcement.” Winter adds that connections are authenticated and encrypted, and that “Identity-driven policies and default-deny network isolation control how agents communicate and what resources they can access.” The instinct is sound. Identity rather than network address becomes the unit of policy, and refusal becomes the default.
The open question is how far down that policy reaches. Letting an agent connect to the HR system is a coarse rule. Restricting it to reading leave balances for one region, without writing, is a fine one, and it is the fine rule that limits damage when an agent is misled. Microsoft’s Ignite description covers connections and resource access. It does not say whether Agent Grid evaluates individual tool calls, methods, or arguments.
Microsoft’s own Container Apps egress engine already reaches host, path, and HTTP method, which is finer than the depth Agent Grid’s Ignite description claims. AWS documents a generally available check on each tool request, and Google’s Agent Gateway now documents tool-level rules without a single release label. A level marked not stated may still be supported once Microsoft publishes Agent Grid documentation, so the exhibit records what is written rather than what the product can do. Until that documentation exists, the fine rule has to live somewhere the enterprise controls.
2What does deeper policy look like in other written designs?
Three written designs. Other written designs show what deeper policy looks like. Microsoft’s own Container Apps egress engine matches rules on host, path, and HTTP method, which is already finer than destination alone. AWS made Policy in AgentCore generally available on March 3, with a gateway that evaluates each agent tool request against Cedar rules before allowing or denying it. Google’s GKE documentation describes a default-deny network policy created for sandboxes built from a template.
What Microsoft’s wording tells a buyer. The quoted lines describe who may talk to whom and what each agent may reach. They do not mention individual tool calls, methods, or arguments. That is not proof the capability is absent, but it marks the boundary of what has been promised. A buyer should read “what resources they can access” as a rule about reach, and should place the rule about action, the one that limits a misled agent, somewhere it can see and test today.
In the source’s own words
- Microsoft’s Book of News, Azure infrastructure foundationAgent Grid will secure interactions among agents, models, tools and Azure services with centralized policy enforcement.Source
- Winter’s Ignite post, on Agent Gridthrough authenticated, encrypted connections.Source
- Winter’s Ignite post, on Agent GridIdentity-driven policies and default-deny network isolation control how agents communicateSource
- Winter’s Ignite post, on Agent Gridand what resources they can access.Source
3Where does Microsoft’s identity record stand?
A record that runs longest. Identity is where Microsoft’s record runs longest. Greyhound Research wrote in May 2025, in Microsoft Build 2025: Making AI Agents More Capable and Secure, that “Entra Agent ID isn’t just a feature,” and called it a policy control plane for agent-based computing. Ignite extends that line. The Book of News says Microsoft “will automatically assign an Agent ID to discovered local agents, starting with GitHub Copilot CLI and Claude Code,” in preview.
What the search found. A search of Microsoft Learn for the Agent Sandbox extension for AKS and for Agent Grid, run on November 18, 2026, up to 02:06 UTC, and repeated on November 19 up to 00:45 UTC, found no Microsoft Learn page for either product.
What remains undocumented. What remains undocumented is the binding between a sandbox session on AKS, the Entra Agent ID it acts under, and the Agent Grid rule that admits its calls.
4What would change if Agent Grid reached the method level?
Two outcomes. Should Agent Grid enforce method-level rules against an Entra Agent ID, Microsoft would have turned privileged access management into a property of the platform, and it would lead on that layer. If its policies stop at the connection, application teams inherit the fine-grained work one agent at a time.
The tests already set. Greyhound Research’s July 2026 note, Everyone Is Building AI Agents. Oracle Is Moving The Contest Inside The Runtime, set tests for any agentic application that include naming each agent’s identity and bounding its authority. Agent Grid will be judged against the same tests.
Part 3Memory And Placement
Context Cache Economics: A Discount On Repetition
Greyhound SpotlightMicrosoft says reused prompt text is billed at a steep discount, and has published no rate. Where other providers publish rates, reused text costs 2.5% to 10% of the normal price.
Context Cache will be sold on price. It saves the model’s work on text an agent sends again and again, so the model does not redo it. Microsoft says this makes long tasks faster and cheaper. The saved work stays in the customer’s own Azure account. Its repository promises a steep discount on cached input and gives no rate.
Caching is not new. Web and database teams have relied on it for decades, and model providers have offered prompt caching for some time. What is new in Microsoft’s version is how long the saved work lasts, which suits agents that pause for a person and resume later. For buyers, the lesson is that the saving belongs to the workload, not the product. Greyhound Research expects the published rate, when it arrives, to matter less than each team’s measured reuse. The second-order effect is on how agents are designed. Once repetition is cheap, engineers have a reason to send longer standing instructions at every step, and that can raise the total bill even as the unit price falls.
Exhibit 5 shows each published price for reused text as a share of the normal input price. A cache read is the cheap reuse of text the model has already processed. Wherever a rate is published, the discount is deep.
Greyhound Research · Exhibit 5 · The price of a cached read
Reused text costs 2.5% to 10% of the normal price
Published prompt-cache read prices as a share of each provider’s base input price, November 19, 2026. Azure has not published its rate.
The discount is deep and narrow. Budget on measured exact repetition, not on the list rate.
Exhibit 5 shows published prices for a cached read as a share of each provider’s base input price, on bars running to 10%, and shows that published cache reads cost 2.5% to 10% of base input while Azure has not published its rate. Anthropic prices cache reads at 0.1 times base input for most models, shown as 10%, at 0.05 times for Opus 5.5 and Sonnet 5.5, shown as 5%, and at 0.025 times for Fable and Mythos 5.1 and 5.2, shown as 2.5%. OpenAI says cached input is discounted up to 95%, shown as 5% at best. Azure Context Cache says cached input is billed at a steep discount and gives no rate, shown as a dashed bar with a question mark. Notes beneath state that only repeated input earns the discount and output is billed in full, that Anthropic charges 1.25 times base input to write a five-minute cache, as OpenAI does for its default cache on GPT-5.6 and later, and that one character of difference in the first 1,024 tokens means nothing is cached.
A deep discount is not the same as a big saving. Output is not discounted, and neither is any input that changes. Writing text into the cache can cost more than normal input. Microsoft has published no rate, so Azure’s bar carries a question mark. A team’s saving depends on how much of its input repeats exactly, and only its own records can show that.
Greyhound Standpoint:
Greyhound Research’s judgment is that Context Cache will save real money for one kind of agent and little for the rest, and Microsoft has not yet told buyers which kind they run. The saving goes to agents that resend a long, unchanging block of text at every step. Agents whose context shifts each turn gain almost nothing, because the discount covers repeated input, not the answer the model writes.
The saving on a bill is the share of input that repeats exactly, multiplied by the discount, less any premium for writing to the cache. Anyone who has run web or database caches knows the rule: a cache pays according to its hit rate, and the hit rate belongs to the workload, not the product. Microsoft is sensible not to publish one savings figure, because no honest one exists, and keeping context for days so agents can pause and resume is an advance rivals do not match today. The limit is that Microsoft has published no rate, so any forecast rests on other providers’ prices.
Before forecasting any saving, FinOps leads should record two weeks of real agent activity and read three numbers from it: how much input repeats exactly, the input share of agent spend, and the write cost once Microsoft publishes it. Budgets belong on measured reuse, not on a list discount.
The Deeper Read · About 4 minutes
Select a question to open the full answer
1How does the cache work, and what exactly has Microsoft written about it?
In Microsoft’s words. Context Cache will be sold on price, so Greyhound Research treats it as a pricing feature first. Winter writes that “Long-context, long-running agents reuse processed prompt context instead of recomputing it on the GPU, so long tasks run faster and cost less,” and that the cached context “stays durable and controlled, by you, in your Azure tenant.” The Azure Context Cache repository describes explicit prompt caching for Azure OpenAI deployments in preview, says “cached input tokens are billed at a steep discount,” and gives no rate.
The depth of the discount is not the size of the saving. Nothing in output, and nothing in input that varies, is discounted, and Anthropic charges 1.25 times base input to write a five-minute cache, and OpenAI’s guide sets the same rate for its default cache write on GPT-5.6 and later. Microsoft’s own prompt caching guide for Foundry says cache writes on GPT-5.6 and later can be charged, with 12-hour cache writes billed from November 1, 2026. For Context Cache, Microsoft has published no rate, so the Azure bar carries a question mark rather than a value. A team’s saving depends on how much of its input repeats exactly, which only its own traces can show.
2Why does caching discount only part of the bill?
Two phases of inference. The mechanism sets the limit. Inference has two phases. In the first, the model processes the entire prompt, and the cost grows with its length. During the second, it generates output one token after another. A cache stores the work of the first phase for reuse. The vLLM documentation states that prefix caching “only reduces the time of processing the queries (the prefilling phase)” and “does not reduce the time of generating new tokens.”
An exact match. Reuse also demands an exact match, and Microsoft’s prompt caching guide says one character of difference in the first 1,024 tokens means nothing is cached.
3What do published prices say?
The ceiling on savings. Published prices set the ceiling on savings. Anthropic’s documentation prices cache reads at 0.1 times the base input price for most models, 0.05 times for Opus 5.5 and Sonnet 5.5, and 0.025 times for Fable and Mythos 5.1 and 5.2, with five-minute cache writes at 1.25 times. OpenAI’s guide says cached input is “discounted up to 95%.” Microsoft has published no rate and no savings figure for Context Cache. A search for Azure Context Cache pricing, with the repository read again, run on November 18, 2026, up to 02:06 UTC, and repeated on November 19 up to 00:45 UTC, found no rate.
4How should a FinOps team calculate the saving?
Three steps. For a FinOps team the calculation runs in three steps. Find the fraction of input tokens that repeat exactly, apply the published read discount to that fraction, then subtract what the cache costs to write. Nothing in output, and nothing in input that varies, is discounted.
Which agents gain. Agents that work over a fixed body of code or policy text see most of their input repeat. Those fielding brief, dissimilar customer exchanges see little, and there a write premium can turn the feature into a net cost.
5What is new in Microsoft’s version?
Durability. Durability is the new element. The Azure CLI reference for context-cache containers sets a time-to-live of 1 to 30 days, counted from the last time the cached context was used, and the repository’s quickstart uses seven. Prompt caches elsewhere typically last minutes, and hours or a day at most on some models.
Built for pause and resume. Agents that pause for a human approval and resume later lose an ordinary cache. Microsoft’s durable version is built for exactly that pattern, and for long-running workflows that is where the larger saving may lie.
What “durable and controlled” tells a buyer. Winter’s words make two promises at once. Durable is the economic promise, since a cache that survives a pause keeps saving. Controlled by the customer is the governance promise, and it carries the customer’s obligations with it. The quickstart’s seven-day setting shows that Microsoft’s own example sits well inside the maximum. Buyers should read it as a starting point to tune for each workload, not as a default to accept.
In the source’s own words
- Winter’s Ignite post, on prompt context reuseLong-context, long-running agents reuse processed prompt context instead of recomputing it on the GPU,Source
- Winter’s Ignite post, on prompt context reuseso long tasks run faster and cost less.Source
- Azure Context Cache repository, quickstarttimeToLive = 7dSource
- Winter’s Ignite post, on prompt context reuseThat cached context stays durable and controlled, by you, in your Azure tenant.Source
Part 3Memory And Placement
Context Cache Governance: A New Store With A Retention Date
Greyhound SpotlightAnything an agent has read, including a password or key someone pasted, can now be kept in processed form for 30 days after its last use, and for as long as it keeps being used.
Context Cache is also a new store of data that most governance programs have not yet listed. Customer data, company code, or a pasted password can all be kept until the cache has gone unused for up to 30 days, so a cache in steady use does not expire. Microsoft keeps it inside the customer’s own Azure account, which is the right place. Only two regions, Central US and Sweden Central, are supported today.
For data protection teams, the cache is a familiar problem in a new place. What differs is that it holds processed context rather than the original text, and no document read yet says what that processed form contains. Whether one person’s record can be found inside it becomes harder to answer. The retention rule also behaves unusually, because a cache in steady use does not expire, so a busy agent can keep material longer than a quiet one. Greyhound Research’s expectation is that many organizations will set one retention value for every workload at first, which is simple to run and hard to defend. Records schedules rarely contemplate a store whose lifetime depends on how often it is used.
Table 2 sets out the entry the cache needs in a data register before its first production use. Microsoft has written four of the fields and not three, and one owner is the customer’s to name.
Greyhound Research · Table 2 · Data register entry
Four fields written, three not stated, one owner to name
The fields a data protection register needs for Azure Context Cache, filled from Microsoft’s own documents, November 19, 2026.
The cache is a data store first. Register it, set retention by workload, and record the three open fields before production.
| Register field | What Microsoft has written | Status | The owner’s decision |
|---|---|---|---|
| Location | Azure Storage resources linked to a model deploymentIn the customer’s tenant, per Winter | Written | Which subscriptions may create caches |
| Retention | Time-to-live of 1 to 30 days, counted from last useDefault 1 day. The quickstart uses 7 | Written | Retention per workload, not the maximum |
| Regions | Central US and Sweden Central onlyLaunch region Central US | Written, preview | Match against residency obligations |
| Provider scope | Explicit prompt caching for Azure OpenAI deployments | Written | Which deployments may cache |
| Contents and form | “processed prompt context”Whether tokens or model state is held is not stated | Not stated | Data classes that must not enter a prompt |
| Price | “billed at a steep discount”No rate given | Not stated | No saving booked before a rate exists |
| Deletion on request | Expiry after disuse onlyNo erasure on request described | Not stated | Test deletion end to end |
| Accountable owner | The customer controls the cache in its tenant | To be named | Name a data owner before first use |
Table 2 sets out Azure Context Cache as a data register entry, and shows four fields written by Microsoft, three not stated, and one owner for the customer to name. Written: the location, Azure Storage resources linked to a model deployment in the customer’s tenant, the retention, a time-to-live of 1 to 30 days counted from last use, with 7 in the quickstart, the regions, Central US and Sweden Central only in preview, and the provider scope, Azure OpenAI deployments. Not stated: the form of the cached contents, the price beyond a steep discount, and a path for erasure on request beyond expiry after disuse. The accountable owner is the customer’s to name. Each row gives the owner’s decision, which is Greyhound Research analysis.
The three missing fields are what the cache holds, what it costs, and how it is erased. Those are the questions a privacy audit asks. Microsoft has written where the cache sits, how long it lasts, its regions, and which models it serves. No document read says what form the saved text takes. Until one does, the register entry should record the gap and name an owner.
Greyhound Standpoint:
Greyhound Research’s stance is that Context Cache belongs in the data inventory before it belongs in the cost plan. A cache that keeps processed context until it has gone unused for up to 30 days is a filing cabinet with a timer, and the timer restarts each time the drawer opens. Customer records, source code, or a pasted password can sit there for a month or longer, in a form no document read yet describes.
Each generation of infrastructure has produced a copy of sensitive data that governance programs found late: backup tapes, replicated databases, log files, and collaboration exports. Each was created for performance or convenience and became an audit finding once someone asked where the data had gone. Erasure requests are usually where such copies surface first. Microsoft deserves credit for keeping the cache inside the customer’s own Azure account, the right home for it, which also makes the customer answerable for it. Table 2 leaves three fields open: what the cache holds, what it costs, and how it is erased. A published deletion path and a description of the stored form would settle most of the concern.
Data protection officers should register Context Cache before any production use. They should set retention by workload rather than at the maximum, check the two supported regions against residency rules, and make an end-to-end deletion test a condition of the pilot sign-off, before any personal data reaches a cache.
The Deeper Read · About 2 minutes
Select a question to open the full answer
1Where is the cache held, for how long, and what is not yet written?
The record in detail. Greyhound Research also reads Context Cache as a data store that most governance programs have not yet catalogued. Anything an agent has read in a prompt, whether customer data, proprietary code, or a credential someone pasted, can now persist in processed form until it has gone unused for up to 30 days. Microsoft’s command-line reference says blobs “that have not been accessed within this number of days will be automatically deleted,” so the clock restarts with each use, and the default is one day. Microsoft holds it inside the customer’s own tenant, and that is the correct home. The repository defines cached context as Azure Storage resources linked to a model deployment, and says only Central US and Sweden Central are supported today.
The three fields Microsoft has not stated, what the cache holds, what it costs, and how it is erased, are the ones a privacy audit tests. Location, retention, regions, and provider scope are written, and the time-to-live of 1 to 30 days is the customer’s to set. The table cannot say what form cached context takes, because no document read says so. Until one does, the register entry should record the gap and name an owner for it.
2What does a data protection team ask of the cache?
The questions of any store. That design raises the questions a data protection team asks of any store. What does the cache contain, and in what form? How long does it live, and who sets that by workload? Which region holds it, and does that region satisfy residency rules? Can it be deleted on request, and does erasure of a customer’s data reach cached prompts derived from it? Copies created for speed and forgotten by their owners are a familiar finding in privacy audits, and an agent cache with a month-long life is a strong candidate.
3Has Greyhound Research made this argument before?
The same argument about memory. Greyhound Research made the same argument about agent memory in August 2025, in Transforming AI Economics with Procedural Memory, whose standpoint held that “without layered governance controls, memory undermines trust faster than it delivers savings.”
A genuine trade-off. Longer retention saves more on agents that pause and resume, while shorter retention limits exposure. No single setting fits all workloads, so someone has to be accountable for choosing.
What the repository header tells a buyer. The header labels the cache a preview with a launch region, and the regions line limits it to two. Read beside Greyhound Research’s earlier warning on memory, those two short lines set the pace for regulated adoption. A company whose residency rules exclude both regions cannot yet use the cache for regulated data, whatever the saving. For everyone else, the preview label means retention and deletion behavior may still change, so register entries should be dated and reviewed at release.
In the source’s own words
- Azure Context Cache repository, headerLaunch region: Central US · PreviewSource
- Greyhound Research, procedural memory notewithout layered governance controls, memory undermines trust faster than it delivers savings.Source
- Azure Context Cache repository, regionsOnly centralus (launch) and swedencentral are supported today.Source
Part 3Memory And Placement
Confidential Computing: The Seal And The Reader
Greyhound SpotlightIgnite’s summary put three confidential computing offerings in preview. Microsoft’s detailed post on them says the rack-scale machines and the Kubernetes containers still “remain in development.”
Ignite’s quietest announcements may matter most to banks, insurers, and health groups. Microsoft’s Ignite summary lists three confidential computing offerings in preview. Confidential computing keeps the cloud provider from reading a customer’s data while it is being processed. The first offering protects very large models running across a whole rack of GPUs. The second puts a smaller sealed space inside a virtual machine, so less has to be trusted. The third is a clean room where several companies can pool data without any of them seeing the others’ records.
Microsoft’s longer Ignite post on confidential computing, by Vikas Bhatia, gives a narrower status. The rack-scale item is a technology showcase, a single-server demonstration with four GPUs. The rack-scale machines, the Kubernetes containers, and the link to Foundry “remain in development.” The smaller sealed spaces are “planned for a future release.” Only the clean room is in preview, and a limited one. What a buyer can sign today is older and narrower: sealed virtual machines as nodes on Microsoft’s Kubernetes service are generally available.
All three keep the cloud operator out. Once released, the rack-scale offer would remove the objection risk committees raise most often about running the largest models in a public cloud. But an agent working inside the sealed space can still read everything it is given. Microsoft’s post names agent memory and retrieved context among the data worth protecting, and does not say that either sits inside the seal. The clean room is also not new, since Microsoft opened its preview in June, and none of the three has a price, region list, or general availability date.
The other question is proof. The hardware signs a statement that it is genuine and running the right code, and auditors will rely on it. Researchers forged that statement in 2025 with under USD 1,000 of equipment and physical access to the server. Google already lets customers use a verifier it does not own. Microsoft’s post says its demonstration checks the proof with a local verifier, and calls remote checking “an architectural direction.” A companion Greyhound Research note on confidential computing across Microsoft, AWS, and Google follows this one.
For regulated buyers, the change is one of scale rather than kind. Confidential computing has been available for years, mostly for single virtual machines, and risk committees have tended to treat it as a specialist control. Ignite points it at the largest models and at agent systems, where those committees now ask their sharpest questions. Greyhound Research expects sealed computing to become a standard line in regulated AI tenders, with the proof behind it examined as closely as the encryption. The clean room carries a quieter consequence. Pooling data no party can see protects the raw records, but what leaves the room is the answer, so the rules on permitted questions decide much of the protection.
Greyhound Standpoint:
Greyhound Research’s view is that Microsoft has shown how much bigger the sealed room can become, and that the harder question is still who sits inside it. Protecting the largest models across a GPU rack would remove the objection risk committees raise most often. Confidential computing seals the envelope against the courier. The agent inside is the reader, and the seal does nothing about what it does with what it reads.
The status has moved since Ignite. Microsoft’s detailed post puts the rack-scale offer and the smaller sealed spaces in development, so Greyhound Research now reads them as a direction to plan for rather than a product to pilot. Microsoft names agent memory as data worth protecting without placing it inside the seal, which risks protecting the arithmetic and leaving the notes on the desk. Its demonstration checks the proof locally, and practitioners learned with hardware security modules that a control is only as strong as the party that checks it. The engineering deserves credit, and the evidence is modest. None of the three has a price, and the 2025 forgery needed physical access that hyperscale sites make much harder. Agent memory documented inside the boundary, with an independent verifier, would strengthen the case.
Before any pilot, regulated security leaders should get written answers on what stays sealed, who verifies the proof and how often, and how a cached prompt is destroyed. Procurement should write those terms into the contract.
The Deeper Read · About 3 minutes
Select a question to open the full answer
1What exactly did Microsoft announce, and how does it compare?
The record in detail. The Ignite 2026 Book of News lists VR72 multi-GPU confidential computing, Azure Confidential Enclaves with AKS Kata CoCo confidential containers, and Azure Confidential Clean Room, and says “Three offerings are now in preview.” A trusted execution environment encrypts data while it is being processed, and the hardware produces a signed statement, known as attestation, that the expected code runs on genuine chips. VR72 stretches that boundary across a rack, Confidential Enclaves shrink it to a smaller box inside a virtual machine, and the clean room shares it between parties.
Two readings of the status. Microsoft’s detailed Ignite post on confidential computing, published by Vikas Bhatia on November 17 and read for this edition on November 19, is narrower than the Book of News. It describes the Ignite item as “A confidential bare-metal technology showcase on NVIDIA Vera Rubin NVL72, with a single-node, four-GPU AI demonstration,” says that “VR72 confidential VMs, AKS and Kata-based confidential containers, and Microsoft Foundry integration remain in development,” and says of the enclave capabilities that “Both capabilities are under active development and planned for a future release.” Only “Azure Confidential Clean Rooms, in limited preview” is available to try. This edition reports both readings and plans on the narrower one. The first edition of this note, which rested on the Book of News, said all three were in preview.
What can be signed today. Microsoft’s AKS Ignite post lists confidential VM node pools with Azure Container Linux, on Intel TDX and AMD SEV-SNP sizes, as generally available. That is protection for a whole node, not the rack-scale or enclave designs announced at Ignite.
The Book of News names the offering “VR72” but not the chip. NVIDIA describes Vera Rubin NVL72 as its first rack-scale confidential computing platform, spanning CPU, GPU, and NVLink. Rivals are moving too. AWS launched Nitro Sandboxes on November 10, 2026, with an optional attested mode, and Google made confidential GPUs and Intel Trust Authority as an independent verifier generally available in June 2026.
2Why does attestation matter more than encryption?
The seal has a lab test. Encryption in use is now standard across the three large clouds, so the difference between offers lies in how the proof is produced, checked, and kept. The TEE.Fail researchers extracted keys from Intel and AMD trusted execution environments and forged attestations, using DDR5 memory bus interposition. Physical access was required, and hyperscale data centers have controls that make this much harder, so this is a lesson about verification rather than a reason to avoid the technology.
Who checks, for now. Microsoft’s detailed post answers part of the question the first edition left open: “Remote attestation is an architectural direction, while the Ignite demonstration uses a local verifier.” Google already offers an independent verifier for its confidential service. A local verifier suits a demonstration. A regulated buyer will want to know who runs the checker in production.
Timing matters too. Attestation proves what was running at one moment. Agents change as models, tools, and prompts are updated, so buyers should ask how often the proof is refreshed, and whether an update forces a fresh check before an agent touches regulated data.
What Microsoft’s own wording tells a buyer. Bhatia’s post is candid on three points that matter for planning. It says which pieces remain in development, it calls remote attestation a direction rather than a feature, and it names agent memory among the data that makes these systems useful. Read together, the lines give a regulated buyer a checklist rather than a verdict: wait for the release terms, ask who will verify the proof in production, and ask where agent memory will sit relative to the seal.
In the source’s own words
- Microsoft, Ignite 2026 Book of Newsprotect large models that extend beyond a single GPUSource
- Microsoft, Ignite 2026 Book of Newswhile helping protect each participant’s sensitive raw data from other collaborators and cloud operators.Source
- AWS Compute Blog, Nitro Sandboxeseach container the boundary of a hardware-isolated virtual machineSource
- Microsoft Community Hub, Bhatia on confidential computing at IgniteVR72 confidential VMs, AKS and Kata-based confidential containers, and Microsoft Foundry integration remain in development.Source
- Microsoft Community Hub, Bhatia on confidential computing at IgniteRemote attestation is an architectural direction, while the Ignite demonstration uses a local verifier.Source
- Microsoft Community Hub, Bhatia on confidential computing at IgniteDatabases hold the business records, retrieved context and agent memory that make those systems useful.Source
Part 3Memory And Placement
AKS AI Manager: Placement Judgment For Scarce GPUs
Greyhound SpotlightNeither of Microsoft’s two Ignite posts on AI Manager mentions KAITO, the open-source project that already does similar work on Kubernetes.
Once AI arrives, the hardest question for a Kubernetes team is where each model should run. AKS AI Manager is Microsoft’s answer. It matches what each model needs to the right chips, called GPUs, across a company’s clusters. Teams can pick from a catalog of models and see cost by deployment.
The shift is from running models to deciding where they run, a judgment many platform teams have not made before. Web and data workloads tolerate a wide range of machines. AI models do not, because memory, throughput, and response time all depend on the chip. A placement tool appeals because it packages scarce expertise. Uptake is likely, in Greyhound Research’s view, to follow the path of managed Kubernetes itself, where convenience won early and questions about portability arrived later. The cost that tends to stay hidden concerns skills. A team that hands placement to a tool learns less about its own workloads, which matters most on the day it wants to move them or challenge a bill.
Exhibit 6 maps the open-source projects Microsoft names near AI Manager and those it does not. The two projects closest to its job, KAITO and llm-d, go unnamed. That matters, because a layer built on shared open-source projects can be moved later.
Greyhound Research · Exhibit 6 · Open-source neighbors
Microsoft names three neighbors, not the closest two
Projects Microsoft names in the Ignite passage that covers AI Manager, projects it does not name, and the preview API’s scope.
Ask what sits beneath AI Manager before adopting it. Open foundations are what keep a serving stack portable.
Exhibit 6 groups the open-source projects around AKS AI Manager into three cards, and shows that Microsoft names three projects near AI Manager and not the two closest to its job. Named by Microsoft in the same Ignite passage: Kubernetes, vLLM, and Ray, without saying they sit beneath AI Manager. Not named by Microsoft at Ignite: KAITO, a CNCF sandbox project accepted on October 17, 2024, and llm-d, which joined the CNCF on March 24, 2026, neither mentioned in the Kubernetes Center post. Context: the preview API, version 2026-09-02-preview, covers bring-your-own clusters and custom models. The relationship between AI Manager and any of these projects is not stated in the documents read.
Greyhound Standpoint:
In Greyhound Research’s view, AI Manager targets the right gap, which is judgment rather than machinery, and its worth depends on how openly that judgment is built. Kubernetes teams know how to run containers. What they struggle with is choosing the chip each model sits on, keeping GPUs busy, and showing finance what each endpoint costs. A seating plan is worth more than another chair.
The decision is costly to get wrong in either direction. Oversize the placement and the company pays for idle silicon. Undersize it and users wait. Cost by deployment turns the conversation with finance from whether AI is expensive to which workload earns its place. Platform veterans have seen how such layers age. Managed layers built over shared open-source projects have tended to stay portable, while proprietary controllers with an open-source surface have tended to become dependencies. Microsoft’s launch material does not name KAITO or llm-d, the two projects closest to this job, so the evidence cannot yet say which path AI Manager is on. Teams strong on Kubernetes and new to GPUs gain most, and mature teams lose nothing if the tool stays optional.
Platform engineering leads should test AI Manager this quarter on one AI service the business can live without, and get written answers on its open-source dependencies, support for clusters outside Azure, cost export to FinOps tools, and how to move a deployment away.
The Deeper Read · About 3 minutes
Select a question to open the full answer
1What does AI Manager do, in Microsoft’s words?
In Microsoft’s words. Kubernetes teams find one question hardest once AI arrives, namely where each model ought to be placed, and Greyhound Research reads AKS AI Manager as an answer to it. Winter writes that it “translates what each model needs into the right infrastructure, placing each deployment where the fleet has the GPU capacity to serve it best.” The Kubernetes Center post says teams can select from a catalog of supported models, deploy to a new or existing cluster, and “view cost by deployment/namespace, and observe inference and GPU metrics.”
2Why is placement hard?
Three variables, one choice. Placement is a hard problem. Memory needs, target throughput, and acceptable latency jointly determine the accelerator type, the count, and the way they are connected. Undersize the allocation and response times suffer. Oversize it and expensive accelerators sit unused. Many platform teams make that choice by hand and revisit it rarely.
Work staffing plans did not assume. Those that built AKS estates for web and data workloads are now asked to broker accelerators, serve model endpoints, and report cost per model, work their staffing plans did not assume.
In the source’s own words
- Winter’s Ignite post, on AKS AI ManagerAKS AI Manager translates what each model needs into the right infrastructure,Source
- Winter’s Ignite post, on AKS AI Managerplacing each deployment where the fleet has the GPU capacity to serve it best.Source
- Microsoft Community Hub, Kubernetes Center postSelect from a catalog of supported models and deploy to a new or existing cluster with your desired SKU and deployment profile.Source
- Microsoft Community Hub, Kubernetes Center postthe AI Manager preview are planned for Microsoft Ignite.Source
3What does the Greyhound CIO Pulse say about the barrier?
Integration complexity. Greyhound Research reported from the Greyhound CIO Pulse 2025 that 64% of enterprises experimenting with GenAI agents cite integration complexity as their top barrier to scaled adoption. A managed placement and catalog layer addresses part of that barrier.
Cost per endpoint. Cost reporting may matter more than placement. Once finance can see spend per endpoint and per namespace, the argument shifts from AI’s cost in general to the return on each workload. Producing that view today usually means hand-joining invoices and cluster telemetry.
4How does AI Manager relate to open source?
Named, and not named. What remains unclear is how AI Manager relates to open source. Winter names Kubernetes, vLLM, and Ray in the passage on flex nodes and an extended commitment to open source, without saying which sit beneath AI Manager. KAITO, the Kubernetes AI Toolchain Operator, already simplifies model inference, tuning, and retrieval workloads on Kubernetes as a CNCF sandbox project accepted on October 17, 2024. llm-d joined the CNCF on March 24, 2026, bringing cache-aware routing and split serving. The Kubernetes Center post does not mention KAITO, and nor does the AKS Ignite post that introduces “AKS AI Manager for Foundry.” That post says “platform teams keep their own networking, identity, policy, and GPU capacity.”
What the names Microsoft chose tell a buyer. Winter names Kubernetes, vLLM, and Ray, all widely used open-source projects, while Microsoft’s posts leave out the two projects closest to AI Manager’s job. The omission may mean little, since launch posts are short. It still leaves the buyer without the fact that decides portability. The AKS post’s line that platform teams keep their own networking, identity, policy, and GPU capacity is encouraging, and a buyer should ask for the same assurance about the placement logic itself.
5What would a closed controller cost?
A shortage traded for a dependency. With a closed controller, the staffing shortage would ease and a supplier dependency would take its place. The AI Manager preview API, versioned 2026-09-02-preview, covers bring-your-own clusters and custom models, which suggests Microsoft expects mixed estates. Mature teams that already run KAITO, KServe, or their own schedulers may not want a second controller, and keeping AI Manager optional for them serves Microsoft’s interest as well as theirs.
Part 4The Customer’s Own Ground
Foundry On Customer AKS: Microsoft’s Catalog On The Customer’s Ground
Greyhound SpotlightMicrosoft says customers will be able to run Foundry’s models and agents on their own Kubernetes clusters. Flex nodes will extend those clusters to servers outside Azure.
For regulated companies, this is the infrastructure announcement with the biggest consequence. Customers will be able to deploy open-weight models, whose trained parameters are published, and agents from Microsoft’s Foundry catalog to their own clusters. Teams can choose chips and regions by cost, speed, and data residency rules.
Regulated buyers gain three things. They decide where data sits by where they place the cluster. The cluster follows their existing network separation. Security reviewers can inspect the actual software before approving it, which a vendor-hosted service does not allow.
This route revives an old enterprise preference in a new form. Regulated firms have long wanted vendor software running on infrastructure they control, from licensed databases to private cloud stacks. What is new is that the software arriving is a catalog of models and agents, which changes faster than any packaged application. Greyhound Research expects the managed and customer-run positions to sit side by side in most large estates. The effect most coverage misses is on the pace of updates. When the catalog moves quickly and the cluster belongs to the customer, someone on the customer’s side has to decide when each new model version is allowed in, and that decision has no published owner yet.
Exhibit 7 places four ways to run Foundry’s catalog on one line, from Microsoft running everything to the customer running everything. The two middle positions have no full written split of duties yet. Each of them moves work from Microsoft to the customer.
Greyhound Research · Exhibit 7 · Who operates the runtime
Four ways to run Foundry’s catalog, two without a full split of duties
Positions on one axis, who operates the runtime, as Microsoft describes them at Ignite 2026, with flex nodes extending the customer end.
The middle positions are the real news. Choose one only once its split of duties is in writing.
Exhibit 7 places four ways to run Foundry’s catalog on one axis, from Microsoft operating the runtime at the left to the customer operating it at the right, labeled Microsoft and The customer, and shows that the two middle positions have no full written division of duties yet. At the Microsoft end is the Foundry managed service, generally available, where Microsoft runs hardware, runtime, and endpoint and the customer sets policy. Next are hosted agents on AKS in managed mode, in preview, named in session BRK420. Then bring-your-own-cluster mode, in limited preview, where the customer operates the cluster and node pools and Microsoft’s catalog, tooling, and governance stay in the control plane. At the customer end is plain AKS with open-source tools. A marker at the middle of the axis shows the seam between the two operators. A note beneath states that flex nodes for AKS, in preview, will extend clusters to customer servers outside Azure.
Greyhound Standpoint:
Greyhound Research concludes that Foundry on customer AKS is the infrastructure announcement with the biggest consequence for regulated enterprises, because it brings Microsoft’s model catalog inside boundaries those enterprises already audit. It ends a false choice between a service Microsoft runs entirely and a stack the customer assembles alone, and it is likely to become the common pattern for regulated AI on every major cloud.
Residency becomes a matter of where the customer places the cluster rather than a negotiation. The model sits behind the network controls that already protect the systems it serves, security reviewers can inspect the runtime they approve, and the customer keeps the logs a regulator will ask for. The cost is easy to leave out. Whoever runs the cluster runs the patches, the capacity, and the incident at three in the morning. The outsourcing waves of earlier decades showed that business cases which priced the license and not the operating staff tended to disappoint once the operating bill arrived. It is also a preview, with no service level yet.
CIOs in banking, healthcare, and the public sector should start a pilot this quarter, with platform engineers and the on-call rota as the first line of the business case. Progress will show in three places: a release date, a published split of duties for each mode, and support for flex nodes outside Azure.
The Deeper Read · About 3 minutes
Select a question to open the full answer
1What exactly did Microsoft announce, and what do regulated buyers gain?
In Microsoft’s words. Greyhound Research regards the Foundry path to customer AKS as the infrastructure announcement with the largest consequence for regulated enterprises this Ignite. The Book of News says customers “will now be able to deploy open-weight models and agents from Foundry to their own Azure Kubernetes Service (AKS) clusters,” and that “Teams will be able to choose GPUs and regions based on cost, latency and data residency.” Winter adds that they move “with consistent infrastructure, security policies, and governance.”
Regulated buyers gain on three counts. Residency is set by where the customer places its cluster, so it no longer depends on what a shared service supports. Network placement follows the enterprise’s existing segmentation. Security reviewers can examine the actual runtime before approving it, something a vendor-hosted endpoint does not permit.
2What did Microsoft describe at Ignite?
The session and the modes. The Ignite session BRK420 describes picking a model in Foundry and deploying it to the customer’s own AKS cluster on GPU node pools, with workload identity, policy, and built-in observability, and running hosted agents on AKS in bring-your-own-cluster and managed modes. Flex nodes for AKS, also in preview, “will extend the same AKS clusters to customer servers outside Azure.” Arun Ulag, Executive Vice President of Microsoft Azure, frames the choice as running “at scale with Azure Kubernetes Service (AKS) when you need more control over performance, security, and cost.” Microsoft’s AKS Ignite post labels both routes, models and hosted agents on the customer’s own clusters, a limited preview, a narrower label than the Book of News gives.
Where regulated pilots stalled. Many regulated pilots stalled at one point: data science wanted the breadth of a managed catalog, and risk wanted every component inside infrastructure it already audits. This path lets both get what they asked for.
What “seamlessly” tells a buyer. Winter’s word describes the deployment path, and the deployment path is the part Microsoft controls. Ulag’s phrasing is more useful for planning, because it presents the customer cluster as the choice for more control over performance, security, and cost, and control arrives with operating work. The limited preview label in the AKS post narrows the claim further. A buyer should read the quotes as a promise about moving software, and plan separately for running it.
In the source’s own words
- Microsoft’s Book of News, Azure infrastructure foundationCustomers will now be able to deploy open-weight models and agents from Foundry to their own Azure Kubernetes Service (AKS) clustersSource
- Winter’s Ignite post, on Foundry to AKSOpen-weight models and agents move seamlessly from Foundry to AKS clusters, deployed from cloud to edge,Source
- Winter’s Ignite post, on Foundry to AKSwith consistent infrastructure, security policies, and governance.Source
- Microsoft’s Book of News, Azure infrastructure foundationFlex nodes for AKS will extend the same AKS clusters to customer servers outside Azure.Source
- Microsoft Azure Blog, Ulag at Igniterun it at scale with Azure Kubernetes Service (AKS) when you need more control over performance, security, and cost.Source
3How does it change the competitive picture?
Both postures from one catalog. Google’s sandbox is a GKE feature, so it lives on clusters the customer operates. AgentCore is AWS’s managed runtime, operated by AWS. Microsoft now covers both postures from a single catalog.
Evidence the customer holds. Evidence is the less obvious gain. Logs, metrics, and network traces from a model on the customer’s cluster belong to the customer, and they are easier to produce for a regulator than records held on a provider’s shared endpoint.
4What comes with the ground?
Operating duties. The caution is that operating duties come with the ground. Node patching, GPU capacity, upgrades, and the overnight page all pass to whoever operates the cluster. Previews carry no service-level commitment.
Capacity moves, and does not disappear. Capacity shifts from model quota on a managed endpoint to virtual-machine quota for specific accelerators in the region the customer chose. Greyhound Research expects every major cloud to make this arrangement, the provider’s catalog and control plane over hardware the customer operates, the standard route for regulated AI within two years.
Staff move with the model. Greyhound Research said in October 2026, of open-weight models run on a customer’s own infrastructure, that “Self-hosting is a staffing decision as much as an infrastructure one, and the people to carry it must move with the model.” The same holds for Foundry on customer AKS.
Part 4The Customer’s Own Ground
Shared Responsibility: The Seam Between Two Operators
Greyhound SpotlightMicrosoft’s Kubernetes team now says the customer runs the cluster and its machines when Foundry agents run on it. Nothing published yet says who answers when a service goes down, or who keeps the logs a regulator may ask for.
The Foundry promise is tested where two operators meet. On Foundry’s managed service, Microsoft runs everything below the customer’s settings. Move it onto a customer’s cluster and the machines and the runtime change hands. The catalog, tools, and governance stay with Microsoft. That split can give convenience and control together. It can also leave a gap that each side expects the other to cover.
For a buyer, the seam is where a pilot’s success can hide a production problem. In a pilot, the same few engineers sit on both sides of the line and close gaps informally. In production, the line runs between two organizations with separate support queues, change calendars, and incentives. Microsoft’s first written split, which places the cluster and its machines with the customer, is a useful and familiar start. The remaining duties are likely to be written gradually, as preview customers raise them. A less obvious consequence lands on support. When an endpoint goes down, the first hours can go on establishing whose queue the ticket belongs in, and the business bears that time.
Table 3 lists the nine duties a split of responsibilities needs, and what Microsoft has written against each. Three are written, four are partly written, and two are not assigned.
Greyhound Research · Table 3 · Responsibility matrix
Nine duties at the seam, two not yet assigned
Duties a responsibility matrix for Foundry on customer AKS needs, by mode, against what Microsoft has written, November 19, 2026.
Sign off only when every row names an owner, for the mode in use. Two rows still have none.
| Duty | Managed mode | Bring-your-own-cluster mode | What Microsoft has written |
|---|---|---|---|
| Policy declaration | Microsoft control plane | Microsoft control plane | Written“consistent infrastructure, security policies, and governance,” per Winter’s post |
| Agent identity | An Entra ID per hosted agent | Not stated for this mode | WrittenFoundry hosted agents documentation |
| Spend controls | Agent 365 Cost Management | Agent 365 Cost Management | Written, previewA centralized place to see and control spend, per the Book of News |
| GPU capacity and quota | Not stated | Customer chooses GPUs and regions | PartlyChoice by cost, latency, and residency, per the Book of News |
| Logs and audit evidence | Not stated | Not stated | Partly“built-in observability,” per BRK420 |
| Policy enforcement on the cluster | Not stated | Customer operates the cluster | PartlyWorkload identity and policy, per BRK420 |
| Node operating system patching | Not stated | Customer operates the node pools | PartlyCustomer runs the cluster and node pools, per the AKS Ignite post |
| Endpoint service level | Not stated | Not stated | Not statedThe path is in preview |
| Model updates and deprecation | Not stated | Not stated | Not stated |
Table 3 lists nine duties for Foundry on customer AKS across managed and bring-your-own-cluster modes, and shows three written, four partly written, and two not yet assigned. Written: policy declaration in Microsoft’s control plane, agent identity as an Entra ID per hosted agent in managed mode, and spend controls through Agent 365 Cost Management in preview. Partly written: GPU capacity and quota, where the customer chooses GPUs and regions, logs and audit evidence, where built-in observability is named, policy enforcement on the cluster, where workload identity and policy are named, and node operating system patching, where the customer operates the node pools in bring-your-own-cluster mode. Not yet assigned: the endpoint service level, and model updates and deprecation. The reading rests on documents published by November 19, 2026, before general availability.
The weak point is the gap between a rule and whether it is enforced. Microsoft has written who sets policy, who gives agents identities, and who controls spend. The service level and model updates have no owner in anything published, and patching falls to the customer only by implication, because it runs the node pools. Auditors who check the console but not the cluster will certify the intention, not the effect. These are previews, so Microsoft may fill some of the empty rows before release.
Greyhound Standpoint:
Greyhound Research finds that convenience and control can coexist on Foundry and AKS, provided the split of duties is written before the first model moves. Shared responsibility is a lease with two signatures, and the signatures matter when something breaks. On a customer cluster the machines and runtime pass to the customer, while Microsoft’s catalog, tools, and governance stay in its control plane, which leaves a seam each side may expect the other to watch.
The weak point is the distance between a rule and its enforcement. A control plane can bar an agent from a database, but enforcement depends on network policy, identity binding, and admission rules on a cluster the customer operates. When the two sit with different parties, an audit must check both, and many check only the first. Early public cloud, and the outsourcing contracts before it, produced the responsibility matrix because unwritten splits turned the first serious incident into a dispute over whose job it was. Microsoft has written three of the nine duties in Table 3 and partly written four, including a first split for hosted agents on customer clusters, a reasonable start for a preview. A complete matrix for each mode would make the hybrid path the sensible default for regulated buyers.
Before production, CISOs should get that split for the mode they choose, walk an incident through it with both teams in the room, and withhold approval until every duty has a named owner.
Part 4The Customer’s Own Ground
Capacity And Quota: The Constraint Behind The Platform Choice
Greyhound SpotlightA new Foundry subscription showed zero capacity for Azure OpenAI models in every region. A moderator on Microsoft’s own Q&A forum called that expected for new customers.
Greyhound Research expects capacity, more than design, to set the pace of Foundry on customers’ own clusters. Quota is the share of scarce AI chips a customer is allowed to use. A new subscription can start at zero, with more requested through a form. Rationing a scarce resource is normal, and Microsoft shares chips across many customers for good reasons.
The order of decisions matters. Residency rules decide the region. The region decides which chips are available, and the chips decide which models can run. A team that picks its region first and asks about capacity second can lose weeks. Greyhound Research’s own reading of Azure capacity has moved since 2025, from the risk of building too much to the limits of supply in a given region, as The Deeper Read explains.
For buyers, the practical change is that capacity becomes a negotiation item rather than an operating detail. That reverses the early promise of public cloud, in which capacity was presented as close to unlimited and bought by the hour. Regional limits sharpen the point, because a residency rule can exclude the region where chips are easiest to find. Greyhound Research anticipates that written capacity in named regions will become part of the commercial conversation with account teams, alongside price. Less discussed is the internal effect. When chips are scarce, the business units that ask first often get the capacity, so a company without its own allocation rule lets timing set its priorities.
Greyhound Standpoint:
For Greyhound Research, quota is the first planning item for any Foundry-on-AKS program, not the last. It is a dull constraint, and the one most likely to move a go-live date. Choosing the platform is the passport. Chip quota in the right region is the visa, and many teams discover the second only at the border.
The mechanism is a chain. Residency rules fix the region, the region fixes which accelerators are available, and the accelerators fix which models a team can run. A new Foundry subscription can begin at zero quota, with more requested through a form. A customer cluster turns that request into one for virtual machine families in one region, and does not remove it. Anyone who planned data centers around power and floor space, or early cloud estates around reserved capacity, has seen scarce supply set the schedule while the architecture debate continued. Rationing scarce chips across many customers is reasonable. The evidence is one onboarding thread that describes standard behavior, not a shortage, so this is a judgment about planning order rather than a forecast of scarcity. Written capacity commitments for named regions would ease it.
Platform leads should secure chip quota in the regions their residency rules allow before designing a single cluster. No production timeline should reach the board until the Microsoft account team has confirmed capacity in writing.
The Deeper Read · About 3 minutes
Select a question to open the full answer
1How does quota work on a managed service and on a customer’s own cluster?
The record in detail. Greyhound Research expects capacity, more than architecture, to set the pace at which enterprises use Foundry on their own clusters. A Microsoft Q&A thread of June 13, 2026, records a new Foundry subscription showing zero quota for Azure OpenAI models in every region, which a moderator described as expected onboarding behavior, with more quota requested through a form. Rationing is standard practice for a scarce resource, and Microsoft shares accelerators across many customers for good reasons.
The dependency runs in one direction. On a managed endpoint, a quota request concerns model throughput. For a customer’s AKS cluster, the request is for particular accelerator-backed virtual machines, in whichever region residency rules dictate. Residency obligations determine the region. What that region can supply in accelerators determines which models are feasible. Global managed endpoints can draw on capacity wherever it exists, and a residency-bound cluster cannot, so a team that settles its region first and asks about capacity second can lose weeks.
2What has Microsoft said about choosing regions and GPUs?
Choice by cost, latency, and residency. For Foundry on customer AKS, the Book of News says teams will be able to choose GPUs and regions based on cost, latency, and data residency. The choice it describes is the customer’s. Written capacity commitments in the regions residency requires are not stated in the documents read, which is why Question 9 asks for them.
What the choice tells a buyer. The sentence gives the customer the choice and is silent on supply. Choosing a region on residency grounds helps only if that region holds the chips the chosen model needs, in the quantity the workload requires. A buyer should read “will be able to choose” as an offer of options, and should ask for the capacity behind each option in writing before treating it as a plan.
In the source’s own words
- Microsoft’s Book of News, Azure infrastructure foundationTeams will be able to choose GPUs and regions based on cost, latency and data residency.Source
3What does a new subscription start with?
Zero, by design. A Microsoft Q&A thread of June 13, 2026, records a new Foundry subscription showing zero TPM quota for Azure OpenAI models in every region. Its moderator described that as expected onboarding behavior, with more quota requested through a form (summarized, not quoted).
A request of a different kind. On a managed endpoint, a quota request concerns model throughput. For a customer’s AKS cluster, it is for particular accelerator-backed virtual machines in whichever region residency rules dictate, so the region is settled before the capacity is known.
4Has Greyhound Research’s view of Azure capacity changed?
From over-build to supply. Yes, and the change deserves stating. Greyhound Research wrote in June 2025 that Microsoft’s AI capital spending plan was “both a statement of strength and a calculated risk,” and that if workloads failed to scale, hyperscalers “could face underutilised infrastructure.” By November 2025 it said that “power availability has overtaken chip availability as the real bottleneck.” Since then, the constraint buyers meet first in planning has been quota for particular chips in particular regions, not idle capacity. The view has therefore moved from whether Microsoft built too much to whether the right accelerators exist where residency rules require them.
What the note does not claim. It does not say Azure as a whole is short of capacity. The evidence here is one onboarding thread describing standard behavior, so the point is about planning order, not a forecast of scarcity.
Part 5The Platform Contest
Three Clouds Compared: Level On Runtime, Ahead On The Estate
Greyhound SpotlightSince March 3, 2026, AWS has checked every tool an agent calls against written rules. Microsoft’s Ignite account of Agent Grid does not claim that depth.
Microsoft calls Azure a single system built together, and Nadella promises an end-to-end trust platform. Both describe integration, which is Microsoft’s strongest suit. A fair comparison with AWS and Google has to go piece by piece, using each vendor’s own documents.
This comparison says less about vendors than about each buyer’s starting point. The integration claim is strongest where a company’s estate already matches it and weakest where it does not. Agents raise the stakes of that match. An agent acts across many systems at once, so the value of an integrated identity and security layer rises with each system it touches, and so does the cost of stitching one together across vendors. AWS and Google are likely to answer with deeper integrations of their own over the coming year. A knock-on effect for buyers is that a single-cloud agent plan and a multi-cloud data estate will pull against each other, and the data estate usually takes longer to move.
Table 4 compares the three, layer by layer. The runtime is the software that runs agents. Here, the estate means everything around it, from identity to office software. Microsoft is level on the runtime and identity, behind AWS on written tool rules, and ahead on the estate.
Greyhound Research · Table 4 · Three clouds, layer by layer
Level on running agents, behind on tool rules, ahead on the estate
Microsoft, AWS, and Google compared layer by layer on each vendor’s own documents, November 19, 2026.
Microsoft leads where the estate is already Microsoft. Score the three against what you run, one layer at a time.
| Layer | Microsoft | AWS | |
|---|---|---|---|
| Isolated runtime | Generally available Container Apps SandboxesAKS extension in preview, VM-level per session | Generally available AgentCore Runtime, a microVM per sessionSince October 13, 2025, next generation November 12, 2026 | Generally available GKE Agent SandboxSince May 20, 2026, gVisor with Kata pluggable |
| Start-up claim | Near-instant from prewarmed poolsThousands concurrent | “under 2 seconds” | 300 per second per cluster90% within 200 milliseconds |
| Agent identity | Documented An Entra ID per hosted agentAgent ID for discovered local agents in preview | Generally available AgentCore IdentitySince October 13, 2025 | Status not on page Agent Identity, SPIFFE-basedDocumentation updated November 17, 2026 |
| Tool-level policy | Preview Agent Grid, identity-driven connectionsTool-call depth not stated | Generally available Policy in AgentCore, Cedar at a gatewayEach tool request, since March 3, 2026 | Not assessed Outside this note’s evidence |
| Network default | Default-deny egress on Container AppsDefault-deny isolation in Agent Grid, preview | Virtual private cloud support at general availability | Default-deny network policy per sandbox template |
| Durable context | Preview Context Cache, 1 to 30 days in the tenant | AgentCore Memory, generally available since October 2025A memory service, not a prompt cache | Not assessed Outside this note’s evidence |
| Model choice | Foundry hosts models from several labs | “any model in or outside Amazon Bedrock” | Not assessed Outside this note’s evidence |
| Customer-run Kubernetes | Preview Foundry on customer AKS, flex nodes | AgentCore is operated by AWS | GKE clusters the customer operates |
| Identity, security, and work estate | Lead Entra, Defender, Purview, and Microsoft 365 beside every agent | Strongest where data and workloads already sit on AWSGreyhound Research reading | Strongest where developers build on Google’s modelsGreyhound Research reading |
Table 4 compares Microsoft, AWS, and Google layer by layer, and shows Microsoft level on runtime and identity, behind AWS on written tool-level policy, and ahead on the estate around the runtime. On runtime and identity, Container Apps Sandboxes, AgentCore Runtime, and GKE Agent Sandbox are generally available, with start-up claims on different bases, and each vendor documents a per-agent identity. On policy and state, AWS’s Policy in AgentCore, generally available since March 3, 2026, checks each tool request at a gateway, while Agent Grid is in preview with tool-call depth not stated. Microsoft’s Context Cache is in preview and AWS offers AgentCore Memory. Microsoft documents running its catalog on customer-run AKS, AgentCore is operated by AWS, and Google’s sandbox runs on customer-operated GKE. Three Google cells are marked not assessed. On the estate around the runtime, Microsoft leads through Entra, Defender, Purview, and Microsoft 365, a Greyhound Research reading.
For a company already standardized on Microsoft 365 and Entra, Microsoft’s identity directory, that head start is valuable and holds up in daily use. Where data lives on AWS, or developers build with Google’s models, the same claim is really a proposal to migrate. Three Google cells are not assessed, so the table does not rank Google on tool rules, saved context, or model choice. On tool rules, AWS is ahead today.
Greyhound Standpoint:
Greyhound Research concludes that Microsoft has earned the claim of the best-integrated agent platform for companies built on Microsoft, not the claim of the only complete one. Every hyperscaler sells a box of runtime, identity, and policy pieces. What separates them is how well the picture on the lid matches the buyer’s estate.
Microsoft is level with AWS and Google on running agents and on identity, behind AWS on written tool rules, and ahead on everything around the agent. An agent built on Foundry inherits Entra, Defender, Purview, and Microsoft 365 from one vendor and saves months of integration a mixed estate does by hand. Where data sits on AWS, or developers build on Google’s models, the same claim reads as a proposal to migrate. Platform contests follow a pattern: features converge within a year or two, as in managed databases and containers, and lasting differences settle in the estate and the operating record. Expect parity on running agents and identity within a year. Within three years, the deciding question will be evidence: whose platform can show an auditor what each agent did, on what authority, and with what result. Three Google cells are not assessed.
Enterprise architects should score the three against the systems they already run at the next architecture review, and pay any integration premium only to the provider that already holds their identities and their most important data.
The Deeper Read · About 2 minutes
Select a question to open the full answer
1What exactly do the vendors claim about integration?
The claims in detail. Greyhound Research tests Microsoft’s platform claim layer by layer rather than as a whole. Winter writes that Microsoft’s work “has forged Azure into a single, co-engineered system,” and Nadella writes, “We are building an end-to-end trust platform.” Both describe integration, and integration is Microsoft’s strongest suit. A fair comparison with AWS and Google has to be made piece by piece, on each vendor’s own documents.
For an organization already standardized on Microsoft 365 and Entra, that head start is valuable, and the integration claim holds up in daily use. Where data lives on AWS, or developers work with Google’s models, the same claim is really a proposal to migrate. Three Google cells are marked not assessed, so the table does not rank Google on tool-level policy, durable context, or model choice. On tool-level policy, AWS is ahead of Microsoft today.
2Where are the three level?
Runtime and identity. On the runtime, the three are level. AWS made Amazon Bedrock AgentCore generally available on October 13, 2025, naming Runtime, Memory, Gateway, Identity, and Observability, and saying it works with “any model in or outside Amazon Bedrock.” GKE Agent Sandbox has been generally available since May, and Google’s Agent Identity documentation, updated November 17, describes “a strongly attested, cryptographic identity for each agent that is based on the SPIFFE standard.” Microsoft’s Entra Agent ID stands alongside both.
3Where does the order change?
Tool-level policy. On tool-level policy, the order changes. AWS’s Cedar-based Policy, generally available since March, inspects every tool request at a gateway, a depth Microsoft’s Ignite account of Agent Grid does not claim. On this layer Microsoft is following.
The estate around the runtime. Its lead sits around the runtime. Entra already holds the identity of nearly every employee in a Microsoft-centered enterprise, Defender already watches those identities, and Microsoft 365 is where much of the work agents will automate begins. Foundry agents plug into that estate without a separate integration project.
4What should a skeptic still credit?
What is real. A skeptic should still credit what is real. Copilot puts Microsoft’s own agents into everyday work, Agent 365 discovery now reaches agents Microsoft did not build, and Microsoft Security treated agents as identities early.
What “co-engineered” tells a buyer. Both quoted phrases describe Microsoft’s real strength, and both describe it from Microsoft’s side. Co-engineered means the parts were built to work together, which saves integration time for a company that uses many of them. An end-to-end trust platform is a promise that reaches beyond what any single document read for this note yet shows. A buyer can test both claims cheaply, by asking which parts of the trust story work for an agent whose data or model sits outside Azure.
Part 5The Platform Contest
Lock-In And Exit: The Agent Carries What The Model Does Not
Greyhound SpotlightSwitching an agent to another AI model is often a settings change. Its permissions, working memory, and history of actions stay with the platform it was built on.
Lock-in now sits in what an agent builds up, not in the model it calls. Changing models has become ordinary, and AWS says its agent service works with any model. Moving the agent itself is harder. Its permissions sit in one vendor’s directory, its memory in that vendor’s storage, and its history in that vendor’s log format. Context Cache is an example, because it is tied to Azure storage and to one model deployment. So is Microsoft’s new Copilot harness for Foundry agents, in preview, which cannot yet be switched once an agent is created.
Lock-in has moved before. In the mainframe era it sat in hardware, later in databases, and in early cloud in proprietary managed services. Each time, buyers negotiated hard on the visible layer and found the binding elsewhere. Agents repeat the pattern, with the model as the visible layer this time. Model portability, which vendors now advertise readily, is therefore a weaker protection than it sounds. Greyhound Research expects export of agent state to become a negotiating point in large contracts before it becomes a published feature. A related effect concerns records. Audit history in one vendor’s format must stay readable for as long as a regulator may ask, which can outlast the platform.
Exhibit 8 splits what an agent depends on by whether it can move to another platform with a settings change today. The model travels, and so do the protocols, the shared standards agents use to talk to tools. Identity, memory, and the record stay.
Greyhound Research · Exhibit 8 · What travels and what stays
The model travels, and identity, memory, and the record stay
What an agent depends on, split by whether it moves to another platform by configuration today, on Microsoft, AWS, and Google alike.
Models move freely and agents do not. Put export of identity, memory, and records into the contract before go-live.
Exhibit 8 splits what an agent depends on into two cards, what travels by configuration and what stays with the platform, and shows that the model and the protocols travel while identity, memory, and the record stay. Travelling today: the model, since Foundry and Bedrock host models from several labs and AWS says AgentCore works with any foundation model, and tool and agent protocols, since MCP and A2A sit with the Agentic AI Foundation. Staying with the platform: identity and permissions, held in each vendor’s directory with no documented export, cached context and memory, since Context Cache is an Azure Storage resource linked to one deployment, also with no documented export, and the audit trail, partly portable because AWS emits OpenTelemetry-compatible telemetry while Microsoft’s export format for agent records is not in the documents read. Each item names what a buyer should get in writing. The split holds for Microsoft, AWS, and Google alike.
After years and hundreds of agents on one cloud, leaving means rebuilding identities, moving saved data, and proving to auditors that the record survived. That is true of Microsoft, AWS, and Google alike. The exhibit does not say any vendor blocks export. It says the documents read describe no route for identity or memory, and only AWS’s audit trail is partly portable.
Greyhound Standpoint:
Greyhound Research’s position is that lock-in has moved from the model to the agent, and that committing to one agent platform is defensible when the way out is in the contract and hard to defend when it exists only in a sales conversation. Models travel light. Agents travel with luggage: the identity they act under, the memory they have built, and the record of what they did, each in its vendor’s format.
The cost of leaving shows up late. After three years and hundreds of agents on any platform, the exit bill lies in recreating identities, migrating state, and proving to auditors that the history survived. Veterans of database and business application migrations know the shape: the license rarely held them, while stored logic, custom data, and audit history did. Open protocols such as MCP and A2A help at the edges, but they carry messages while memory stays where it was made. Building in-house is no answer either, since Microsoft’s tools bring identity and monitoring that would take years to match. The documents read describe no export route for identity or memory on any cloud, a finding about documents, not a claim that any vendor blocks export.
CIOs should use Microsoft’s agent tools where the integration earns its keep. Before any agent goes live, they should make export of definitions, permissions, saved state, and audit records a signed term of the contract.
The Deeper Read · About 3 minutes
Select a question to open the full answer
1Where exactly does each piece of an agent live?
The detail. Greyhound Research locates agent lock-in in what an agent accumulates, not in the model it calls. Changing models has become ordinary. Foundry and Bedrock each offer models from multiple labs, AWS’s documentation says AgentCore works “with any foundation model,” and repointing an agent to another model is frequently a settings change. Moving the agent itself is harder. Its rights are recorded in a directory, its working context sits in a storage service, and its history of actions is kept in a log schema, each belonging to the platform it was built on. Context Cache is an example, because Microsoft provisions it as an Azure Storage resource bound to one model deployment.
After several years and hundreds of agents on any one cloud, the price of departure is rebuilding identities, moving state, and showing auditors that the record came through intact. That holds for Microsoft, AWS, and Google equally. The exhibit does not say that any vendor blocks export, only that the documents read describe no route for identity or memory. AWS’s OpenTelemetry-compatible telemetry leaves the audit trail as the one item partly portable.
2Where do open standards help, and where do they stop?
MCP and A2A. Standards ease the boundaries. The Agentic AI Foundation, part of the Linux Foundation, now hosts both MCP and A2A, and it announced on August 17 that A2A was joining as a hosted project, and AWS names both protocols in its AgentCore launch. They give agents a common way to invoke tools and exchange messages. What they leave in place is an agent’s memory and its record of decisions.
In the source’s own words
- AWS Documentation, AgentCore overviewwith any foundation modelSource
3Is this consistent with what Greyhound Research has said before?
The open agentic web. This keeps faith with what Greyhound Research wrote in May 2025 in Microsoft Build 2025: Supporting the Open Agentic Web, that “the agentic web shouldn’t be vendor-controlled or locked behind opaque APIs.” Model choice on Foundry is open, and nothing in this research note describes it as closed. The binding sits in identity, state, and evidence. Greyhound Research said the same in April 2026, as Microsoft and OpenAI changed their contract terms: “Lock-in is not going away. It is relocating.” and “At the model level, substitution is becoming easier.” This note names where it has relocated to.
What the earlier positions tell a buyer. Greyhound Research’s view has held steady: open model choice is welcome, and it does not settle portability. The April 2026 line that lock-in is relocating rather than disappearing is the practical one for buyers. It means the negotiation should follow the binding to where it now sits, in identity, saved state, and records, rather than concentrate on model access, which vendors now offer readily.
4Is a self-built stack the answer?
Vendor reliance for scarce engineers. Retreating to a self-built stack would be a mistake. An Entra and Microsoft 365 customer gets identity, security monitoring, and agent discovery from Microsoft that an internal team would need years to replicate, and Microsoft’s discovery now extends to local tools such as GitHub Copilot CLI and Claude Code. Building it all internally exchanges reliance on a vendor for reliance on scarce engineers.
The outage note’s reasoning. Greyhound Research’s November 15 outage note held that consolidating on one suite remains right for most enterprises once the terms around it are strengthened, and the same reasoning holds for agents.
Part 5The Platform Contest
Buy Or Build: Buy The Commodity, Own The Decisions
Greyhound SpotlightContainer Apps sandboxes cost nothing in processor or memory fees while stopped, and Foundry’s hosted agents bill only during active sessions. Convenience is now priced by the hour.
The line between buying and building should follow what a company will one day have to prove. Ignite gives a range of options rather than a choice of two. Microsoft can run everything. Its catalog can run on the customer’s cluster. Or the customer can assemble open-source parts itself. The middle option matters most, because a team can start on Microsoft’s service and move later without rebuilding its agents.
The rule is simple. Buy what is standard and backed by its provider: the machinery that runs agents, the sandbox, and access to models. Keep what expresses the company’s own judgment: what each agent may do, where a person must approve, and the tests that decide whether an agent’s work is acceptable. Regulators and courts will examine the second group, and it is also the hardest to move.
The range of options is itself the news. For most of the history of enterprise software, buyers chose between a vendor’s hosted service and a stack they assembled, and moving between the two meant a rebuild. Microsoft’s middle path lowers that switching cost, at least on paper. For buyers, that turns buy or build from a one-time decision into one that can be revisited layer by layer as needs change. On Greyhound Research’s reading, the decision layer, rather than the runtime, is where most companies will find their skills gap. The effect that is easy to miss is organizational. Owning the decision layer means staffing policy and evaluation work that rarely appears in a platform budget.
Greyhound Standpoint:
Greyhound Research maintains that the buy-or-build line should follow what a company will later have to prove, not engineering taste. Buy what is standard and provider-backed, such as the runtime, the sandbox, and the model catalog. Keep the layers that express how the business decides: which tools an agent may call, which actions need a person, and how outcomes are judged. Directors will ask who set an agent’s limits, and if that layer was bought, the answer sits in a supplier’s product.
Business software taught this before. Companies that heavily customized their core business applications often lost a clean upgrade path, while those that kept the core standard moved faster. Agent platforms add a twist, because the core is young and this year’s leader may trail next year, which argues for buying with an exit in hand. Microsoft’s middle option, its catalog on the customer’s own cluster, makes that exit cheaper than it was, and it deserves credit. Managed services price convenience into every hour, while self-run clusters price it into salaries and idle capacity, so building the runtime pays only for steady, high demand and a mature team.
By March 31, 2027, CIOs should put the agent estate on one page that marks each layer as bought or owned. They should buy with a rehearsed exit and sign nothing unless definitions, saved state, and decision history can be extracted within ninety days.
The Deeper Read · About 2 minutes
Select a question to open the full answer
1What are the layers, and how should each be treated?
The rule in detail. Greyhound Research draws the buy-or-build line according to what the enterprise will one day have to prove. Ignite leaves enterprises with a graduated set of options rather than a binary one. Microsoft can operate everything, Microsoft’s catalog can run on the customer’s cluster, or the customer can assemble open-source components on AKS itself. The intermediate option is the significant one, because a team can begin on the managed service and relocate later without rebuilding its agents.
The rule that holds is simple to state. Purchase what is undifferentiated and certified by its provider, which today means the execution environment, the isolation layer, and access to models. Retain what expresses the enterprise’s own judgment: the permissions each agent holds, the points at which a person must approve, and the tests that decide whether an agent’s output is acceptable. Regulators and courts will examine the second group, and it is also the hardest to migrate.
2When does owning the runtime pay off?
Narrow cases. Owning the runtime pays off in narrow cases. Models no managed service supports, sovereignty obligations a shared service cannot meet, or accelerator demand high and constant enough to undercut managed pricing are the usual ones, and each assumes a mature platform team.
Cost enters the rule. Cost has to enter the rule too. A managed service charges for convenience by the hour. Self-operated clusters charge for it through payroll and through hardware that waits for work.
3What pattern should large enterprises expect?
Priced for bursts. The Container Apps documentation notes, “You pay no CPU or memory fees when sandboxes are stopped,” which suits bursty agent work, and Foundry’s hosted agents bill on CPU and memory during active sessions.
A mixed estate. Greyhound Research expects the large-enterprise pattern within twenty-four months to be mixed: managed agents where work is general, customer-operated clusters where inference is regulated or heavy, and an enterprise-owned layer of policy and evaluation spanning both. A balanced portfolio passes one test: its policy layer can serve a second platform unchanged.
What the billing lines tell a buyer. Both quotes describe pricing that rewards intermittent work, which fits most agent pilots well. The same pricing changes shape at volume. An agent estate that runs steadily through the day pays for convenience in every active hour, and that is when the case for moving some workloads to the customer’s own cluster begins to strengthen. A buyer should model both patterns before committing, using its own measured activity rather than the pilot’s.
Part 6The Risks And The Counter-Argument
Risks To Watch: Where Agent Infrastructure Can Go Wrong
Greyhound SpotlightNone of the six risks starts inside a single product. Each sits where products meet, and each shows first in the buyer’s own estate.
Greyhound Research sees six risks in the new infrastructure, most of them where products meet. Each shows first inside the buyer’s own organization, where it can be caught early.
The first three concern the platform. Previews can change before release, and AI Manager’s interface still carries a preview version label, 2026-09-02-preview. If Agent Grid governs connections but not tool calls, fine-grained limits fall back on each application. Rights granted to make a demonstration work tend to stay. The last three concern the buyer’s assumptions. A forecast built on a list discount rather than measured reuse will miss. Keeping a cache at maximum retention, or in a region that breaks residency rules, is a privacy finding waiting to happen. On a customer cluster, an incident with no owner becomes an argument, and quota shortfalls show up as missed dates.
None of these risks is specific to agents, and that is the useful point for buyers. Previews that change, rights that outlive a demonstration, forecasts built on list prices, and duties nobody owns have appeared in every platform shift. What agents change is speed and number. A company that once added a workload every few weeks may add agents daily, often from different teams. The first problems are likely to surface in ordinary control processes, such as access reviews, budget variances, and privacy assessments, rather than as dramatic incidents. A shift in reporting tends to follow. Platform teams that once reported uptime will be asked to report agent rights and spend as well.
Exhibit 9 sets each risk beside its early warning, its owner, and the Runbook move that answers it. Every early warning can be checked inside the buyer’s own organization.
Greyhound Research · Exhibit 9 · Risks by early warning and owner
Every risk shows first in the buyer’s own estate
Six risks in the new agent infrastructure, each with its early warning, its owner, and the Runbook move that answers it.
Every risk here is cheaper to prevent than to undo, and each early warning can be checked this quarter without Microsoft.
Exhibit 9 sets out six risks in the new agent infrastructure as cards, each with its early warning, its owner, and the Runbook move that answers it, and shows that every early warning can be checked by the buyer without help from Microsoft. Risk one, preview drift, warns through a production design that references a preview interface, owned by platform engineering. Risk two, policy that stops at the connection, warns through tool limits hand-coded agent by agent, owned by the chief information security officer. Risk three, privilege carried out of the pilot, warns through an agent with no named owner or written tool list, owned by the chief information security officer. Risk four, savings booked from a list price, warns through a budget line quoting a percentage without a hit rate, owned by the FinOps lead. Risk five, a cache nobody catalogued, warns through a cache with no register entry, owned by the data protection officer. Risk six, a dispute at the seam, warns through a responsibility matrix with empty rows, owned by the chief information officer and procurement. The risks are Greyhound Research analysis.
The six share a cause: an assumption that the platform will supply a control the buyer did not specify. Each warning sits in something the buyer already holds, such as a design document, an agent register, a budget line, or a responsibility matrix. The exhibit names owners and moves, not likelihoods, and ranks none of the risks above another.
Greyhound Standpoint:
In Greyhound Research’s reading, none of the six risks is exotic, and each is cheaper to prevent than to undo. They share a cause: an assumption that the platform will supply a control the buyer did not specify. A design that relies on a preview, an agent without an owner, or a cache without a register entry is quick to fix before production and slow to fix after it, once the gap has users and exceptions of its own.
Many of the costliest problems in earlier platform shifts came from seams between products, from integration assumptions nobody wrote down and access granted for a demonstration and then kept. Agent infrastructure concentrates those seams, since sandboxes, gateways, caches, and identities often belong to different teams. The register has a limit. It names owners and early warnings, not likelihoods, and it ranks none of the risks above another, because the evidence does not support a ranking. Documented Agent Grid rules at the tool call, a deletion path for the cache, and a duty matrix for each Foundry mode would each ease one of the six.
Buyers can check each early warning this quarter without help from Microsoft, and each has an owner in the Runbook that follows. Risk committees should take the six at their next quarterly review. This is a judgment about order, not a prediction of harm.
Part 6The Risks And The Counter-Argument
The Strongest Objection: The First Wave Only Needs A Sandbox
Greyhound SpotlightMicrosoft already runs more than a million sandboxes a day. The strongest objection to this note starts there, and most of it stands.
The strongest objection deserves its full weight. Today’s agents in production are mostly coding helpers, analysis assistants, and internal automation. Their main hazard is running code nobody checked. A quick, throwaway machine for each session removes that hazard, and Microsoft already runs more than a million sandboxes a day. On this view, controls on what agents may do can wait, and Ignite delivered what the first wave needs.
The objection persuades because it matches what most companies run today, and ordering security work by present risk is sound practice. Its weakness is that agent estates change character without a clear signal. A coding helper becomes a workflow agent when it gains one more tool, and nothing in the sandbox marks the change. Earlier automation waves drifted the same way, as convenience scripts became systems the business depended on. Greyhound Research anticipates that the move from read-only agents to agents that change records will arrive gradually, one permission at a time, rather than as a planned second wave. That makes the grant of a new tool, not the release of a new product, the moment to review authority.
Exhibit 10 tests each part of the objection against the evidence. Three parts stand. The claim that controls on authority can wait falls short on timing.
Greyhound Research · Exhibit 10 · The objection held to the evidence
The objection holds for code execution and falls short on timing
The strongest case against this research note, limb by limb, against the evidence gathered for each.
The objection holds for code, not for timing. Buy the sandbox first, and design the authority layer at the same pilot gate.
| The objection claims | What the evidence shows | Verdict |
|---|---|---|
| Today’s production agents are mostly software helpers, analysis assistants, and internal automation | That matches the first wave this research note describes. | Stands |
| Their chief hazard is executing code nobody vetted | For the first wave, a sandbox is the largest single reduction in risk. | Stands |
| Microsoft already operates sandboxes at scale | More than a million a day across Microsoft, and Container Apps Sandboxes is generally available. | Stands |
| Authority controls are a later problem | Coding agents already hold repository write access and pipeline credentials, and workflow agents already read mail and documents, which is where EchoLeak sat. Permissions granted in the first wave become the defaults of the second. | Falls on timing |
Exhibit 10 holds the four limbs of the strongest objection to this research note against the evidence, and shows that three limbs stand and the claim that authority controls are a later problem falls on timing. The claims that today’s production agents are mostly software helpers, analysis assistants, and internal automation, that their chief hazard is executing unvetted code, and that Microsoft already operates sandboxes at more than a million a day all stand. The claim that authority controls are a later problem falls on timing, because coding agents already hold repository write access and pipeline credentials, workflow agents already read mail and documents, where EchoLeak sat, and permissions granted in the first wave become the defaults of the second. The verdicts are Greyhound Research analysis.
Greyhound Standpoint:
Greyhound Research accepts the objection for running code and does not accept it for the rest of what agents now do. Today’s production agents are mostly coding helpers, analysis assistants, and internal automation, and their main hazard is running code nobody checked. A fresh, disposable machine for each session removes that hazard, and Microsoft’s volume of more than a million sandboxes a day shows the design holds under real load. A sandbox is the right first purchase, and three parts of the objection stand.
The part that does not hold is timing. The claim that controls on authority can wait assumes rights are as easy to narrow later as they are to grant now. Experience points the other way. Identity programs have spent years removing standing access and shared service accounts granted for speed during earlier projects, and each removal met a business owner ready to argue that something would break. The objection would gain ground if most agents stayed read-only for the next year, or if Agent Grid arrived with tool-level rules that could be applied to running agents without redesign. The evidence available assures neither.
Platform and security leads should buy the sandbox now and design the authority layer alongside it, at the same pilot gate, because narrowing the rights of agents already in production is slower and more contested than granting narrow rights from the start.
The Deeper Read · About 1 minute
Select a question to open the full answer
1Why does the objection fall short on timing?
The first wave already holds access. For the first wave, a sandbox is the largest single reduction in risk, and the generally available Container Apps product is the strongest case for adopting now. The objection falls short on timing rather than substance. Coding agents already hold repository write access and pipeline credentials, and internal workflow agents already read mail and documents, which is where EchoLeak sat.
In the source’s own words
- Microsoft Learn, Container Apps Sandboxes overviewEach sandbox runs in its own secure boundarySource
2What happens when the second wave arrives?
The second wave inherits the first. The second wave, agents that change records in finance, HR, and customer systems, arrives on the same infrastructure, and permissions granted in the first wave become the defaults of the second.
What “overprivileged” tells a buyer. Gallot’s word describes agents already deployed, not a future risk. If Microsoft’s own security leadership finds internally deployed agents frequently hold more access than they need, the first wave already carries the rights the objection says can wait. That makes a review of current agent permissions the cheapest place to start, before any new infrastructure is bought.
In the source’s own words
- The Official Microsoft Blog, Gallot on October 19These agents are frequently overprivilegedSource
Part 7What To Do, By Role
The Greyhound CXO Runbook: Ten Moves, None Waiting For Microsoft
Greyhound SpotlightEvery move uses a control the buyer already holds: its pilots, its identity governance, its gateways, its usage records, its data register, its contracts, or its buying calendar.
Ten moves follow from the evidence, at least one for each leader who will own part of the new infrastructure. Each names the assumption it corrects, the evidence behind it, and the executive who decides. None waits for Microsoft, because each uses controls the buyer already holds.
The moves are built around decisions the buyer already makes, and that is deliberate. Pilots, access reviews, data registers, budgets, and contract renewals all run on calendars a company controls. Attaching agent controls to those calendars makes them hard to skip and cheap to audit, where a separate program would compete for staff. Earlier platform shifts produced many governance programs that drafted policies and then stalled once ownership blurred. The first versions will be rough, and should improve as Microsoft publishes terms. A point that is easy to overlook is sequence. Several moves lean on an accurate agent register, so the registration work in the second move sets the pace for the rest.
Table 5 sets the ten moves beside the assumption each corrects, its evidence, and its owner. Two of the ten finish only when Microsoft publishes its terms.
Greyhound Research · Table 5 · The Greyhound CXO Runbook by owner
Ten moves, and none waits on Microsoft to start
Ten moves for the leaders who will own part of the new agent infrastructure, with the assumption each corrects, its evidence, and its owner.
Start all ten this quarter. Only two need Microsoft’s published terms to finish, and neither is a reason to wait to begin.
Select a move to open its full text in The Deeper Read below
| Move | The assumption it corrects | Evidence anchor | The decision sits with |
|---|---|---|---|
| 1/Pilot on the generally available product, and keep previews out of production | A keynote announcement is a production commitment | Table 1, Exhibit 2 | Head of platform engineering, with the CIO |
| 2/Register every agent as a privileged identity before it receives a tool | A sandbox contains agent risk | Exhibit 3 | CISO, with the identity lead |
| 3/Hold tool-level limits in a gateway the enterprise controls until Agent Grid is documented | Connection policy also governs the tool call | Exhibit 4 | Chief security architect, with the CISO |
| 4/Measure reuse before forecasting any saving | A published discount is a budget saving | Exhibit 5 | FinOps lead, with the CFO |
| 5/Put Context Cache in the data register before the first production cache exists | A cache is a performance setting | Table 2 | Data protection officer |
| 6/Test AI Manager on one inference service the business can live without | A managed placement layer is portable by default | Exhibit 6 | Head of platform engineering |
| 7/Start the regulated pilot with the staffing line, not the price line | The business case is a price comparison with a managed endpoint | Exhibit 7 | CIOs in banking, healthcare, and the public sector |
| 8/Withhold approval until every duty at the seam has an owner | A promise of consistent policy assigns the duties beneath it | Table 3 | CISO |
| 9/Draw the buy-build map and contract the exit before the first agent goes live | Lock-in lives in the model | Exhibit 8, Table 3 | CIO, with the CISO |
| 10/Sign on general availability terms, not on keynote dates | Preview pricing and availability will hold at launch | Table 1, Capacity And Quota | Head of procurement, with the CIO |
Table 5 sets out the ten moves of the Greyhound CXO Runbook, each with the assumption it corrects, the exhibit or table it rests on, and the executive with whom the decision sits, and shows that none waits on Microsoft to start while two complete only when Microsoft publishes its terms. The moves are: pilot on the generally available product and keep previews out of production, register every agent as a privileged identity, hold tool-level limits in a gateway the enterprise controls, measure reuse before forecasting any saving, put Context Cache in the data register, test AI Manager on one inference service, start the regulated pilot with the staffing line, withhold approval until every duty at the seam has an owner, draw the buy-build map and contract the exit, and sign on general availability terms. The moves are Greyhound Research analysis.
The two that finish later, on savings and on signing, wait on prices and terms Microsoft has not yet published. Those terms are what the questions that follow ask for, and each question maps to a move a buyer can make meanwhile. Every other move rests on an exhibit or table in this note and names who decides. The remaining eight can be finished with controls the buyer already holds.
The Deeper Read · About 4 minutes
Select a move to open its full text
1Platform engineering leads: pilot on the generally available product, and keep previews out of production.
Table 1 and the sandbox section show one generally available sandbox product and a preview extension with no published documentation, region list, or price. The move corrects the habit of treating a keynote announcement as a production commitment. The move is to run this quarter’s pilot on Container Apps Sandboxes or the AKS preview using non-production data, with go-live conditions written in advance: a documented Entra Agent ID for every session, a service level in writing, and a dated general availability for Agent Grid. Request GPU quota in the regions residency allows before designing any cluster. The decision sits with the head of platform engineering, together with the CIO.
Why the first move comes first. Pilots set precedents. The identities, rights, and conditions an agent receives during a pilot tend to become the template for production, so writing go-live conditions before the pilot starts shapes every later move. Requesting quota at the same time is a scheduling choice rather than a technical one, since capacity in the right region is the item most likely to move a date.
2CISOs: register every agent as a privileged identity before it receives a tool.
Exhibit 3 shows that isolation answers one route of harm in three. The move corrects the belief that a sandbox contains agent risk. The move is to give every agent an accountable owner, an enumerated set of permitted tools, approval gates where actions cannot be reversed, and a quarterly review that withdraws unused rights. The decision sits with the CISO, together with the identity lead.
3Security architects: hold tool-level limits in a gateway the enterprise controls until Agent Grid is documented.
Exhibit 4 and the Agent Grid section show a policy described at the connection, with no Microsoft Learn documentation through the evidence cut-off. The move corrects the assumption that connection policy also governs the tool call. The move is to ask Microsoft in writing for the finest unit an Agent Grid policy can name, how a session binds to an identity, and where credentials are injected, and until those answers exist to keep tool-level restrictions in the application or in a gateway the enterprise controls. The decision sits with the chief security architect, together with the CISO.
4FinOps leads: measure reuse before forecasting any saving.
Exhibit 5 shows that published cache discounts apply only to repeated input, and Microsoft has published no rate for Context Cache. The move tests the assumption that a published discount is a budget saving. The move is to capture two weeks of production agent traces, measure the exact-repeat share of input and the input share of spend, and book savings only from those measurements once Microsoft publishes its read and write prices. The decision sits with the FinOps lead, together with the CFO.
5Data protection officers: put Context Cache in the data register before the first production cache exists.
Table 2 shows a store that holds processed prompts until they have gone unused for up to 30 days, in two preview regions, with expiry after disuse but no written path for erasure on request. Caches are often treated as performance settings outside the data inventory, and this move corrects that. The move is to register the cache, set retention per workload rather than at the maximum, check the regions against residency obligations, and test deletion end to end before personal data reaches it. The decision sits with the data protection officer.
6Platform engineering leads: test AI Manager on one inference service the business can live without.
Exhibit 6 shows that Microsoft names three open-source projects near AI Manager and not the two closest to its job. The move tests the assumption that a managed placement layer is portable by default. The move is to run the test this quarter, and to get written answers on AI Manager’s open-source dependencies, its support for clusters beyond Azure, the export of its cost data to existing FinOps tooling, and the steps to migrate a deployment away. The decision sits with the head of platform engineering.
7Regulated CIOs: start the Foundry-on-AKS pilot with the staffing line, not the price line.
Exhibit 7 shows four ways to run Foundry’s catalog, with operating duties passing to whoever runs the cluster. The move corrects the habit of building the business case as a price comparison with a managed endpoint. The move is to start a pilot this quarter, with the platform engineers and on-call rota it requires as the first line of the business case, and to track three signs of progress: a date for general availability, a published division of duties for each mode, and support for flex-node clusters beyond Azure. The decision sits with CIOs in banking, healthcare, and the public sector.
8CISOs: withhold approval until every duty at the seam has an owner.
Table 3 shows nine duties at the seam, two of them not yet assigned in anything Microsoft has published and four only partly written. The move corrects the assumption that a promise of consistent policy assigns the duties beneath it. The move is to obtain from Microsoft a duty-by-duty allocation for whichever mode is chosen, walk an incident through it with both teams in the room, and withhold approval until no duty is left without an owner. The decision sits with the CISO.
9CIOs: draw the buy-build map and contract the exit before the first agent goes live.
Exhibit 8, Table 3, and the sections on Foundry on customer AKS, lock-in, and buy or build show that agent lock-in sits in identity, state, and evidence, and that duties at the seam are not yet written. The move replaces the belief that lock-in lives in the model. The move is to draw a one-page map by the end of March 2027, own the decision layers, require a written division of duties for whichever Foundry-on-AKS mode is chosen, and write export of definitions, permissions, cached state, and audit records into the contract. The decision sits with the CIO, together with the CISO.
10Procurement and vendor management: sign on general availability terms, not on keynote dates.
Table 1, the preview standpoint, and the quota section show previews with no published price or service level, and capacity that depends on region. Under test is the assumption that preview pricing and availability will hold at launch. The move is to hold any commitment of budget beyond 2027 until general availability terms are published, require capacity commitments in writing, and compare offers again after AWS’s re:Invent, which opens on November 30. The decision sits with the head of procurement, together with the CIO.
Part 7What To Do, By Role
Questions For Microsoft: What Buyers Will Ask Next
Greyhound SpotlightThree of the ten questions have no answer at all in what Microsoft has published, including what leaves with the customer.
Every finding in this note rests on what Microsoft has published, and Microsoft alone can close every gap it names. Ten questions follow, in the order the evidence raises them. Each gives what Microsoft has said so far, what the answer would settle, and the Runbook move it maps to. No move waits on them, and an account team can answer each from documents Microsoft already holds. The questions are numbered Q1 to Q10 so answers can be tracked as they arrive.
These questions are best asked now, while the products are in preview and their terms are still being drafted. In a preview, customer feedback can still shape documentation, service levels, and contract language before they settle. Once a product ships, its terms tend to be offered as standard and changed only at renewal. The ten also differ in kind. Some ask Microsoft to publish a fact it already holds, such as a price or a region list. Others ask for a commitment, such as written capacity or an export route. Greyhound Research expects the factual questions to be answered before the commitments. A written answer filed with the pilot decision is worth more than a spoken one.
Table 6 tracks the ten questions against what Microsoft has said so far. Seven are answered in part, and three not at all.
Greyhound Research · Table 6 · Questions for Microsoft, tracked
Seven questions answered in part, three not at all
The ten questions only Microsoft can answer, against what it had stated by 00:45 UTC on November 19, 2026.
Ask in writing while the terms are still forming. No Runbook move waits on these answers to start.
Select a question to open its full text in The Deeper Read below
| Question | What Microsoft has said so far | The Runbook move it settles |
|---|---|---|
| Q1Where do the new sandboxes run, and how do the three sandbox products relate? | Stated in part The written Ignite record places the Agent Sandbox extension on AKS, with a serverless option on Azure Container Instances. ACI Sandboxes was shown at Build as an emerging technology preview. | Runbook move 1 |
| Q2What service level, region list, price, and general availability date apply to the AKS extension and Agent Grid? | Not stated No Microsoft Learn documentation for either product was found through November 19, 2026, 00:45 UTC. | Runbook moves 1 and 10 |
| Q3How far down does Agent Grid’s policy reach? | Stated in part Identity-driven, default-deny connection policy, and control of what resources agents can access. | Runbook move 3 |
| Q4How is each sandbox session bound to an Entra Agent ID, and where are credentials injected? | Stated in part Each Foundry hosted agent gets its own Entra ID. Container Apps egress keeps credentials outside the sandbox. | Runbook move 2 |
| Q5What does Context Cache cost? | Stated in part Cached input tokens are “billed at a steep discount”. No rate is given. | Runbook move 4 |
| Q6What does Context Cache hold, and how does erasure reach it? | Stated in part Azure Storage resources linked to a model deployment, 1 to 30 days, in Central US and Sweden Central. | Runbook move 5 |
| Q7Which open-source projects sit beneath AKS AI Manager? | Stated in part Winter names Kubernetes, vLLM, and Ray nearby. The Kubernetes Center post does not mention KAITO. | Runbook move 6 |
| Q8Who does what in each Foundry-on-AKS mode? | Stated in part Two modes are named, with a promise of consistent policy and governance, and the customer operates the cluster for hosted agents. | Runbook moves 7 and 8 |
| Q9What capacity can Microsoft commit in writing in the regions residency requires? | Not stated Teams choose GPUs and regions. Quota for a new subscription can start at zero. | Runbook moves 1 and 10 |
| Q10What leaves with the customer? | Not stated No export format for agent definitions, permissions, cached context, or audit records in the documents read. | Runbook move 9 |
Table 6 tracks ten questions only Microsoft can answer against what it had stated by 00:45 UTC on November 19, 2026, and shows seven answered in part and three not stated at all. Answered in part: where the new sandboxes run, how far down Agent Grid’s policy reaches, how a session binds to an Entra Agent ID, what Context Cache costs, what it holds, which open-source projects sit beneath AI Manager, and who does what in each Foundry-on-AKS mode. Not stated: the service level, regions, price, and general availability date for the AKS extension and Agent Grid, written capacity commitments in the regions residency requires, and what leaves with the customer. Each question names the Runbook move it settles. The grading is Greyhound Research analysis.
Two answers would change this note’s reading most. One is a documented Agent Grid rule that reaches each tool call and is tied to an agent’s identity. The other is a published split of duties for each way of running Foundry on a customer’s cluster. Either would move Microsoft from following to leading on the layer this note calls the control. None of the ten is a trick question. Each asks Microsoft to write down what its product does or will do at full release.
The Deeper Read · About 4 minutes
Select a question to open its full text
Q1Where do the new sandboxes run, and how do the three sandbox products relate?
Said so far. Microsoft’s written Ignite record places the Agent Sandbox extension on AKS, Container Apps Sandboxes is generally available, and ACI Sandboxes was shown at Build as an emerging technology preview. The answer settles. The host, the isolation technology, and the basis of each start-up figure for each product. Evidence that would settle it. Product documentation for each, with its status. Greyhound Research would update. Exhibit 2 and Table 1.
Q2What service level, region list, price, and general availability date apply to the AKS extension and Agent Grid?
Said so far. Not stated. The Book of News says all capabilities are in preview except Cobalt 200, and searches of Microsoft Learn through the evidence cut-off found no documentation for either product. The answer settles. When the go-live conditions of the first Runbook move, and the signing terms of the tenth, can be met. Evidence that would settle it. Published documentation and general availability terms. Greyhound Research would update. Table 1.
Q3How far down does Agent Grid’s policy reach?
Said so far. Microsoft describes identity-driven, default-deny connection policy and control of what resources agents can access. The answer settles. Whether a rule can name an individual tool call, method, or argument, and so whether the third Runbook move can lean on the platform or must be built in each application. Evidence that would settle it. Agent Grid documentation naming the finest unit a policy can address. Greyhound Research would update. Exhibit 4 and Table 4.
Why the depth of the rule matters. This question decides where much of the buyer’s engineering effort goes. If Agent Grid can name a tool call, the fine rules can sit in one platform layer. If it cannot, each application team has to build and maintain its own, and the buyer pays for that work many times over. A single page of documentation would settle which, and that is why it leads the signals in the Final Greyhound Standpoint.
Q4How is each sandbox session bound to an Entra Agent ID, and where are credentials injected?
Said so far. Each Foundry hosted agent gets its own Entra ID, and Container Apps egress keeps credentials outside the sandbox. Neither is documented for the AKS extension. The answer settles. Whether the agent register of the second Runbook move can rely on platform identity for AKS sessions. Evidence that would settle it. Documentation of the session-to-identity binding and of credential injection outside the sandbox, as on Container Apps. Greyhound Research would update. Exhibit 1.
Q5What does Context Cache cost?
Said so far. The repository promises a steep discount on cached input and gives no rate. The answer settles. The read price, any write premium or storage charge, and so the saving the fourth Runbook move can book. Evidence that would settle it. A published price list. Greyhound Research would update. Exhibit 5.
Q6What does Context Cache hold, and how does erasure reach it?
Said so far. Cached context is held as Azure Storage resources linked to a model deployment, for 1 to 30 days, in Central US and Sweden Central. The answer settles. The form in which context is stored, how erasure requests reach it, and which regions follow Central US and Sweden Central, which together decide the register entry of the fifth Runbook move. Evidence that would settle it. Documentation of contents, deletion, and the region roadmap. Greyhound Research would update. Table 2.
Q7Which open-source projects sit beneath AKS AI Manager?
Said so far. Winter names Kubernetes, vLLM, and Ray near AI Manager, and the Kubernetes Center post does not mention KAITO. The answer settles. Whether AI Manager builds on KAITO or llm-d, whether it runs on clusters outside Azure, and whether its cost data can flow into FinOps tools from other vendors, which decides how portable a serving stack remains under the sixth Runbook move. Evidence that would settle it. An architecture statement for AI Manager. Greyhound Research would update. Exhibit 6.
Q8Who does what in each Foundry-on-AKS mode?
Said so far. Microsoft promises consistent policy and governance across managed and bring-your-own-cluster modes, and its AKS Ignite post says the customer operates the cluster, node pools, networking, registry, and storage for hosted agents. The answer settles. The owner of node patching, the endpoint service level, logs and audit evidence, policy enforcement on the cluster, and model updates, for each mode, which the seventh and eighth Runbook moves need. Evidence that would settle it. A written responsibility matrix for each mode. Greyhound Research would update. Table 3 and Exhibit 7.
Q9What capacity can Microsoft commit in writing in the regions residency requires?
Said so far. Not stated. The Book of News says teams will be able to choose GPUs and regions, and a Microsoft Q&A thread records a new Foundry subscription starting with zero quota in every region. The answer settles. Whether a residency-bound plan can hold its dates, under the first and tenth Runbook moves. Evidence that would settle it. Written capacity commitments from the Microsoft account team. Greyhound Research would update. The reading in Capacity And Quota.
Q10What leaves with the customer?
Said so far. Not stated. No export format for agent definitions, permissions, cached context, or audit records appears in the documents read. The answer settles. Export formats for agent definitions and permissions, a documented route for extracting cached context and memory, and audit records in an open telemetry standard, which the ninth Runbook move writes into the contract. Evidence that would settle it. Published export documentation. Greyhound Research would update. Exhibit 8.
Part 8The Final Greyhound Standpoint
Azure Agent Infrastructure: The Final Greyhound Standpoint
Greyhound SpotlightFive signals would move this position, and Microsoft controls all five.
Greyhound Research’s position is that Ignite 2026 made running agents on Azure ready to pilot, confirmed the estate around them as Microsoft’s real advantage, and left the layer that decides what an agent may do still to be written. Once agents have their own machines, the machine stops being the question. What each agent may read, change, and spend becomes the control, and proof that it stayed inside those limits becomes the record buyers will be asked to show.
On running agents, Microsoft is level with AWS and Google. Container Apps Sandboxes is generally available and proven at Microsoft’s own scale, while the Kubernetes extension is a preview without documentation, regions, or price. On written tool rules, AWS is ahead today, Google documents comparable rules without a single release label, and Agent Grid’s depth will decide whether Microsoft closes the gap or leads. The estate is where Microsoft is ahead. Entra, Defender, Purview, and Microsoft 365 give a Foundry agent identity, monitoring, and data protection that would take rivals and internal teams years to build.
Authority becomes the control for economic reasons as much as technical ones. Earlier platform shifts, from server virtualization to public cloud, followed a similar arc: the new layer was sold on speed, became a commodity within a few years, and lasting cost and risk settled with whoever governed access, data, and the exit. Agents are moving along that arc faster, because they act on the business directly rather than through a person.
For regulated buyers, Foundry on a customer’s own cluster carries the most weight of the week’s news, because it brings Microsoft’s catalog inside boundaries they already audit, provided the duties are written down and the staff are in place. Context Cache is a real saving for agents that reuse the same text, a modest one for the rest, and a new data store for everyone. Confidential computing points to a much larger sealed room, most of it still in development, without settling what an agent does inside it. AI Manager fills a real gap, and its openness will decide how easily teams can move. On every cloud, lock-in now sits in what an agent builds up, not in the model it calls.
The evidence has limits, and they set how hard to push. Most of what Microsoft announced is in preview. Agent Grid and the Kubernetes extension had no documentation when the evidence was read, and an absent document is not an absent capability. Start-up speeds are company claims measured in different ways, and statuses, regions, and prices are as published through November 19, 2026.
Five signals would move this position. A documented Agent Grid rule at the tool call, tied to an agent identity, would put Microsoft in the lead on authority. Four more would strengthen the case for production: a duty-by-duty split for each Foundry mode, a service level and release date for the Kubernetes extension, a published cache price with a deletion path, and documented export of agent identity, state, and audit records.
Exhibit 11 sets those signals side by side with the party that controls each. All five sit with Microsoft, which largely sets the pace at which this position can change.
Greyhound Research · Exhibit 11 · Signals that would move the position
Five signals would move this position
What would move Greyhound Research’s judgment, and the party that controls each, with the one event that could reset the comparison.
A documented tool-level rule in Agent Grid would move this view furthest. Watch its documentation first.
Exhibit 11 sets out in three cards the signals that would move Greyhound Research’s position, with the party that controls each, and shows that all five signals sit with Microsoft. One signal would put Microsoft in the lead on authority: a documented Agent Grid policy at the individual tool call against an Entra Agent ID. Four would strengthen the case for production commitment: a duty-by-duty allocation for each Foundry-on-AKS mode, a service level and general availability date for the AKS extension, a published Context Cache price with a deletion path, and documented export of agent identity, state, and audit evidence. One event controlled by AWS could reset the comparison: agent runtime news at re:Invent, which opens on November 30, 2026. The signals are Greyhound Research analysis.
One event outside Microsoft’s control could still reset the comparison: news about running agents at AWS’s re:Invent on November 30. Until the signals arrive, CIOs and platform leads should pilot now on what is generally available, CISOs and data protection officers should own the decision layers and registers from the first agent, and procurement should sign only on full-release terms, revisiting each commitment at the first renewal after release.
Part 9Appendix, Sources, And Disclosures
Appendix: The Evidence Register, Source By Source
Greyhound Research sets out here the evidence behind every fact, figure, and quotation in this research note, one drop-down for each class of source, so that any sentence can be checked against the words it rests on without leaving the page. Each row gives the source, when it was read, the part of the source the words sit in, what it says, and where in this research note that material is used. Times of reading are in UTC. For this edition, every source except four news reports was read again between 00:10 and 00:45 UTC on November 19, 2026, after Microsoft’s longer Ignite posts on Kubernetes and on confidential computing had become readable. Microsoft’s security update page for CVE-2025-32711 did not render on that pass, so its score was checked against the published CVE record, in which Microsoft rates it 9.3 and the US National Vulnerability Database 7.5, and its row keeps the earlier reading time. The time shown is the time of the latest reading. Where a passage was summarized and not quoted, the row says so. A passage quoted in The Deeper Read of a section is not repeated here, and the source’s entry names the section that quotes it.
Microsoft announcements and blogs14 sources, 58 extracts
| Source, read, and used | Where in the source | What it says |
|---|---|---|
| Bhatia, V. (2026)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Confidential Computing, Azure Agent InfrastructureQuoted in The Deeper Read of Confidential Computing | VR72 | [quoted in The Deeper Read] |
| Status | [quoted in The Deeper Read] | |
| Enclaves | [quoted in The Deeper Read] | |
| Clean rooms | [quoted in The Deeper Read] | |
| Attestation | [quoted in The Deeper Read] | |
| Data to protect | [quoted in The Deeper Read] | |
| Hollan, J. et al. (2026)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Lock-In And Exit | Status | The harness is introduced “in public preview” (quoted fragment). |
| Preview limits | “Existing agents are unchanged, switching the harness after creation isn’t supported yet” | |
| Billing | “Billing for the Copilot harness in Foundry Agent Service begins December 1, 2026” | |
| Microsoft (2026a)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in What Ignite Changed, What Microsoft Announced, Agent Grid And Identity, Foundry On Customer AKS, Shared Responsibility, Capacity And Quota, Three Clouds Compared, Lock-In And Exit, Buy Or Build, Questions For Microsoft. Exhibit 7. Tables 1 and 3Quoted in The Deeper Read of What Microsoft Announced, Agent Grid And Identity, Foundry On Customer AKS, Capacity And Quota, Shared Responsibility | Azure infrastructure foundation for production AI applications and agents | [quoted in The Deeper Read] |
| Same heading | [quoted in The Deeper Read] | |
| Same heading | [quoted in The Deeper Read] | |
| Same heading | [quoted in The Deeper Read] | |
| Same heading | [quoted in The Deeper Read] | |
| Same heading | [quoted in The Deeper Read] | |
| Azure Cobalt 200 | [quoted in The Deeper Read] | |
| Agent 365 Cost Management | [quoted in The Deeper Read] | |
| Microsoft Entra | [quoted in The Deeper Read] | |
| Foundry, continuous improvement | Foundry Agent Service is listed among "Foundry capabilities that are generally available" (summarized, not quoted). | |
| Microsoft (2026i)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Confidential Computing | Post | Announces the preview of multiparty analytics in Azure Confidential Clean Rooms (summarized, not quoted). |
| Microsoft (2026j)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in What Microsoft Announced, Confidential Computing, AKS AI Manager, Foundry On Customer AKS, Shared Responsibility, Questions For Microsoft. Tables 1, 3, and 6Quoted in The Deeper Read of What Microsoft Announced, AKS AI Manager, Shared Responsibility | Isolated sandboxes | [quoted in The Deeper Read] |
| Isolated sandboxes | [quoted in The Deeper Read] and [quoted in The Deeper Read] | |
| Foundry hosted agents | [quoted in The Deeper Read] | |
| Headings | [quoted in The Deeper Read] | |
| AKS AI Manager for Foundry | [quoted in The Deeper Read] | |
| Confidential computing | Confidential VM node pools with Azure Container Linux, on Intel TDX and AMD SEV-SNP sizes, are generally available (summarized, not quoted). | |
| Reading note | The post does not mention KAITO and gives no service-level commitment. | |
| Winter, J. (2026)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in What Ignite Changed, What Microsoft Announced, Sandboxes At Scale, Agent Grid And Identity, Context Cache Economics, Context Cache Governance, AKS AI Manager, Foundry On Customer AKS, Shared Responsibility, Three Clouds Compared, Questions For Microsoft. Exhibits 1, 2, 4, and 6. Table 1Quoted in The Deeper Read of Sandboxes At Scale, Agent Grid And Identity, Context Cache Economics, AKS AI Manager, Foundry On Customer AKS, Three Clouds Compared | Agent Sandbox extension | [quoted in The Deeper Read] |
| Agent Sandbox extension | [quoted in The Deeper Read] and [quoted in The Deeper Read] | |
| Agent Grid | [quoted in The Deeper Read] […] [quoted in The Deeper Read] […] [quoted in The Deeper Read] | |
| Prompt context reuse | [quoted in The Deeper Read] […] [quoted in The Deeper Read] | |
| Prompt context reuse | [quoted in The Deeper Read] | |
| AKS AI Manager | [quoted in The Deeper Read] […] [quoted in The Deeper Read] | |
| Foundry to AKS | [quoted in The Deeper Read] […] [quoted in The Deeper Read] | |
| Open source | [quoted in The Deeper Read] | |
| Closing | [quoted in The Deeper Read] | |
| Reading note | The post does not use the words Context Cache, micro, ACI, Container Instances, thousands, or half a second. | |
| Ulag, A. (2026)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Foundry On Customer AKSQuoted in The Deeper Read of Foundry On Customer AKS | Post | [quoted in The Deeper Read] |
| Nadella, S. (2026)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in What Microsoft Announced, Shared Responsibility, Three Clouds Compared. Table 1Quoted in The Deeper Read of What Microsoft Announced, Three Clouds Compared, Shared Responsibility | Post | [quoted in The Deeper Read] |
| Post | [quoted in The Deeper Read] | |
| Agent 365 | [quoted in The Deeper Read] | |
| Reading note | The post does not use the word insider. | |
| Borkar, R. and Kishan, A. (2026)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in What Microsoft Announced. Table 1Quoted in The Deeper Read of What Microsoft Announced | Post | [quoted in The Deeper Read] |
| Post | [quoted in The Deeper Read] | |
| Post | [quoted in The Deeper Read] | |
| Post | [quoted in The Deeper Read] | |
| Customer quote | Amadeus’s Luc Choubert reports a benchmark gain of more than 30% in flight search throughput over Cobalt 100 (quoted in the text of The Deeper Read). | |
| Holeček, A. (2026)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in What Ignite Changed, Agents As Insiders, Three Clouds ComparedQuoted in The Deeper Read of Agents As Insiders, What Ignite Changed | Post | [quoted in The Deeper Read] |
| Dynamic Agent Risk | [quoted in The Deeper Read] | |
| Post | [quoted in The Deeper Read] | |
| Gallot, H. (2026)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Agents As Insiders, The Strongest ObjectionQuoted in The Deeper Read of Agents As Insiders | Post | [quoted in The Deeper Read] |
| Post | [quoted in The Deeper Read] | |
| Hulme, M. (2026)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in What Microsoft Announced, Sandboxes At Scale, The Strongest Objection. Exhibits 1 and 10 | Post | More than a million sandboxes a day run in production across Microsoft (summarized, not quoted). |
| Microsoft (2026b)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in What Microsoft Announced, Questions For Microsoft. Table 1Quoted in The Deeper Read of What Microsoft Announced | Page description | [quoted in The Deeper Read], demonstrated at Build. |
| Nair, H. (2026)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in What Microsoft Announced, AKS AI Manager, Questions For Microsoft. Exhibit 6Quoted in The Deeper Read of AKS AI Manager | AI Manager | [quoted in The Deeper Read] |
| AI Manager | [quoted in The Deeper Read] | |
| AI Manager | [quoted in The Deeper Read] | |
| Reading note | The post does not mention KAITO. |
Microsoft documentation, repositories and sessions10 sources, 29 extracts
| Source, read, and used | Where in the source | What it says |
|---|---|---|
| Microsoft (2026c)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in What Ignite Changed, What Microsoft Announced, Sandboxes At Scale, Buy Or Build, The Strongest Objection, Questions For Microsoft, Azure Agent Infrastructure. Exhibit 2. Tables 1 and 4Quoted in The Deeper Read of What Microsoft Announced, The Strongest Objection, Buy Or Build | Overview | [quoted in The Deeper Read] |
| Overview | [quoted in The Deeper Read] | |
| Overview | [quoted in The Deeper Read] | |
| Billing | [quoted in The Deeper Read] | |
| Reading note | The page does not mention ACI or AKS. | |
| Microsoft (2026d)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Agents As Insiders, Agent Grid And Identity, Questions For Microsoft. Exhibit 4Quoted in The Deeper Read of Agents As Insiders | Recommended posture | [quoted in The Deeper Read] |
| Transform and Rewrite | [quoted in The Deeper Read] | |
| Rules | Rules match on host, path, and HTTP method (summarized, not quoted). | |
| Reading note | The page shows a publication date of May 21. Its update metadata is inconsistent. | |
| Microsoft (2026e)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in What Microsoft Announced, Agents As Insiders, Shared Responsibility, Buy Or Build, Questions For Microsoft. Tables 1 and 3Quoted in The Deeper Read of What Microsoft Announced, Agents As Insiders, Buy Or Build | Isolation | [quoted in The Deeper Read] |
| Identity | [quoted in The Deeper Read] | |
| Billing | [quoted in The Deeper Read] | |
| Regions | Thirty-one regions are listed (counted, not quoted). | |
| Reading note | The page carries no status label. | |
| Microsoft (2026f)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Context Cache Economics. Exhibit 5 | Requirements | A minimum of 1,024 tokens, and one character of difference in the first 1,024 tokens means nothing is cached (summarized, not quoted). |
| Retention | An extended retention of 24 hours is described (summarized, not quoted). | |
| Cache writes | On GPT-5.6 and later, cache writes can be charged, with 12-hour cache writes billed from November 1, 2026 (summarized, not quoted). | |
| Microsoft (2026g)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Foundry On Customer AKS, Shared Responsibility. Table 3. Exhibit 7 | Session description | A model is picked in Foundry and deployed to the customer's own AKS cluster on GPU node pools, with workload identity, policy, and built-in observability, and hosted agents run on AKS in bring-your-own-cluster and managed modes (summarized, not quoted). |
| Microsoft (2026h)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Capacity And Quota, Questions For Microsoft. Table 5 | Question and answer | A new Foundry subscription shows zero TPM quota in every region. A moderator describes it as expected onboarding behavior, with more quota requested through a form (summarized, not quoted). |
| Microsoft (n.d.a)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in What Microsoft Announced, Context Cache Economics, Context Cache Governance, Lock-In And Exit, Questions For Microsoft. Exhibit 5. Tables 1 and 2Quoted in The Deeper Read of Context Cache Governance, Context Cache Economics | Header | [quoted in The Deeper Read] |
| Pricing | [quoted in The Deeper Read] | |
| Regions | [quoted in The Deeper Read] | |
| Quickstart | [quoted in The Deeper Read] | |
| Reading note | The repository gives no rate. | |
| Microsoft (n.d.b)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Context Cache Economics, Context Cache Governance, Questions For Microsoft. Table 2 | Parameters | The time-to-live parameter accepts 1 to 30 days (summarized, not quoted). |
| Time-to-live | [quoted in The Deeper Read] | |
| Microsoft (n.d.c)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in AKS AI Manager, Risks To Watch. Exhibit 6 | Pull request title | The API version, 2026-09-02-preview, and its scope, bring-your-own clusters and custom models, are taken from the title (summarized, not quoted). |
| Microsoft Security Response Center (2025)Read November 18, 2026, between 16:25 and 16:40 UTCUsed in Agents As Insiders, The Strongest Objection. Exhibits 3 and 10 | Security update record | CVE-2025-32711, M365 Copilot Information Disclosure Vulnerability, released June 11, 2025, with a CVSS base score of 9.3 (summarized, not quoted). |
| FAQ | Fully mitigated by Microsoft, with no action for users of the service to take (summarized, not quoted). |
AWS and Google11 sources, 19 extracts
| Source, read, and used | Where in the source | What it says |
|---|---|---|
| Amazon Web Services (2025)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Three Clouds Compared, Lock-In And Exit. Table 4Quoted in The Deeper Read of Three Clouds Compared | Announcement | Runtime, Memory, Gateway, Identity, and Observability are named, with A2A and MCP (summarized, not quoted). |
| Announcement | [quoted in The Deeper Read] | |
| Amazon Web Services (2026a)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Agent Grid And Identity, Three Clouds Compared, Azure Agent Infrastructure. Exhibit 4. Table 4 | Announcement | Cedar policies at a gateway evaluate each agent tool request (summarized, not quoted). |
| Amazon Web Services (2026c)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Confidential ComputingQuoted in The Deeper Read of Confidential Computing | Introduction | [quoted in The Deeper Read] |
| Confidential mode | An optional attested mode releases secrets only to workloads running code the customer approved, and the service is broadly available (summarized, not quoted). | |
| Amazon Web Services (n.d.)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Foundry On Customer AKS, Lock-In And Exit. Exhibit 8Quoted in The Deeper Read of Lock-In And Exit | Overview | [quoted in The Deeper Read] |
| Observability | Telemetry is OpenTelemetry-compatible (summarized, not quoted). | |
| Google Cloud (2026c)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Confidential Computing | Confidential Space | Confidential Space support for NVIDIA Hopper GPUs is generally available (summarized, not quoted). |
| Attestation | Intel Trust Authority is generally available as an independent attestation verifier for Confidential Space (summarized, not quoted). | |
| Parthasarathy, M. (2026)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Sandboxes At Scale. Exhibit 2. Table 4Quoted in The Deeper Read of Sandboxes At Scale | Post | [quoted in The Deeper Read] |
| Post | [quoted in The Deeper Read] | |
| Amazon Web Services (2026b)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Sandboxes At Scale, The Greyhound CXO Runbook, Azure Agent Infrastructure. Exhibit 11 | Event page | The event runs from November 30 to December 4, 2026 (summarized, not quoted). |
| Royal, B. and Hockin, T. (2026)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Sandboxes At Scale, Three Clouds Compared. Exhibit 2. Table 4Quoted in The Deeper Read of Sandboxes At Scale | Post | [quoted in The Deeper Read] |
| Post | [quoted in The Deeper Read] | |
| Post | gVisor isolation and a default-deny network policy, with Kata pluggable (summarized, not quoted). | |
| Google Cloud (2026b)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Agent Grid And Identity, Foundry On Customer AKS. Exhibit 4. Table 4 | Network policy | A default-deny NetworkPolicy is created for each SandboxTemplate (summarized, not quoted). |
| Google Cloud (2026a)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Three Clouds Compared. Table 4. Exhibit 8Quoted in The Deeper Read of Three Clouds Compared | Overview | [quoted in The Deeper Read] |
| Reading note | The page carries no general availability or preview label. | |
| Google Cloud (2026d)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Agent Grid And Identity, Azure Agent Infrastructure. Exhibit 4 | MCP traffic | The gateway extracts tool names and method attributes from MCP requests, enabling tool-level access policies. Connections are blocked by default unless an IAM policy grants access. The page, last updated November 18, 2026, carries no single release label, and Cloud Run and cross-cloud use are in limited preview (summarized, not quoted). |
Model providers, open source, security research, and media17 sources, 22 extracts
| Source, read, and used | Where in the source | What it says |
|---|---|---|
| Anthropic (n.d.)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Context Cache Economics. Exhibit 5 | Pricing | Cache reads at 0.1 times base input, 0.05 times for Opus 5.5 and Sonnet 5.5, and 0.025 times for Fable and Mythos 5.1 and 5.2. Five-minute cache writes at 1.25 times (summarized, not quoted). |
| NVIDIA (2026)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Confidential Computing | Engineered to Scale Intelligence | Vera Rubin NVL72 is the first rack-scale platform to deliver NVIDIA Confidential Computing, across CPU, GPU and NVLink domains (summarized, not quoted). |
| OpenAI (n.d.)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Context Cache Economics. Exhibit 5Quoted in The Deeper Read of Context Cache Economics | Pricing | [quoted in The Deeper Read] |
| Cache writes | On GPT-5.6 and later, the default cache write costs 1.25 times the uncached input rate (summarized, not quoted). | |
| TEE.fail researchers (n.d.)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Confidential Computing | Research page | DDR5 memory bus interposition is used to extract keys from Intel TDX and AMD SEV-SNP and to forge attestations, with a setup estimated at under $1,000 (summarized, not quoted). |
| Toulas, B. (2025)Read November 18, 2026, for the first editionConsulted for Confidential Computing. Cited in the text through the researchers’ own page | Article | The attack required physical access to the server (summarized, not quoted). |
| vLLM (n.d.)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Context Cache EconomicsQuoted in The Deeper Read of Context Cache Economics | Limits | [quoted in The Deeper Read] |
| Limits | [quoted in The Deeper Read] | |
| Cloud Native Computing Foundation (n.d.)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in AKS AI Manager. Exhibit 6Quoted in The Deeper Read of AKS AI Manager | Project page | A sandbox project, accepted on October 17, 2024 (summarized, not quoted). |
| Project page | [quoted in The Deeper Read] | |
| Cloud Native Computing Foundation (2026)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in AKS AI Manager. Exhibit 6 | Post | Split prefill and decode serving, and prefix-cache-aware routing (summarized, not quoted). |
| Agentic AI Foundation (2026)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Lock-In And Exit. Exhibit 8 | Post | A2A joins the Agentic AI Foundation as a hosted project, alongside MCP (summarized, not quoted). |
| Cloudflare (n.d.)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Sandboxes At Scale | Product page | Isolated sandboxes for running untrusted code, with AI agent code execution listed as the first use (summarized, not quoted). |
| Amure, T. O. (2026)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Agents As InsidersQuoted in The Deeper Read of Agents As Insiders | Article | The essay title, "Models as Insider Risks in the Super Intelligence Era." |
| Article | [quoted in The Deeper Read] | |
| Lakshmanan, R. (2025)Read November 17 or 18, 2026, before 02:00 UTC on November 18Consulted for Agents As Insiders and The Strongest Objection. Cited in the text through Microsoft Security Response Center (2025) | Article | CVE-2025-32711, rated 9.3 on CVSS, fixed by Microsoft on the server side (summarized, not quoted). |
| Help Net Security (2026)Read again November 19, 2026, between 00:10 and 00:45 UTCConsulted for Agents As Insiders. Not used in the text | Article | Excessive Agency moves up to third place in the OWASP 2026 Top 10 for LLM applications (summarized, not quoted). |
| Infosecurity Magazine (n.d.)Read November 17 or 18, 2026, before 02:00 UTC on November 18Consulted for Agents As Insiders. Cited in the text through Chismon (2025) | Article | The UK National Cyber Security Centre warns that these systems draw no inherent distinction between data and instruction (summarized, not quoted). |
| Chismon, D. (2025)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Agents As InsidersQuoted in The Deeper Read of Agents As Insiders | Blog post | [quoted in The Deeper Read] |
| Riley, D. (2026)Read again November 19, 2026, between 00:10 and 00:45 UTCConsulted for Agents As Insiders. Cited in the text through Gallot (2026) | Article | Gallot, executive vice president of Microsoft Security, is reported as naming the "agentic insider" (summarized, not quoted). |
| Article | Enterprise agents are described as often overprivileged and growing in number (summarized, not quoted). | |
| The New Stack (2026)Read November 17 or 18, 2026, before 02:00 UTC on November 18Consulted for Sandboxes At Scale. Cited in the text through the vendors’ own pages | Headline | AWS, Google Cloud, Microsoft Azure, and Cloudflare all offer agent sandboxes, each built differently (summarized, not quoted). |
Greyhound Research notes and published comments10 sources, 13 extracts
| Source, read, and used | Where in the source | What it says |
|---|---|---|
| Greyhound Research (2025a)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Sandboxes At Scale, Agent Grid And Identity, Three Clouds ComparedQuoted in The Deeper Read of Sandboxes At Scale, Agent Grid And Identity | Note | [quoted in The Deeper Read] |
| Note | [quoted in The Deeper Read] | |
| Greyhound Research (2025b)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in AKS AI Manager, Lock-In And ExitQuoted in The Deeper Read of AKS AI Manager, Lock-In And Exit | Note | Sixty-four percent of enterprises experimenting with GenAI agents cite [quoted in The Deeper Read], from the Greyhound CIO Pulse 2025. |
| Note | [quoted in The Deeper Read] | |
| Greyhound Research (2025c)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Context Cache GovernanceQuoted in The Deeper Read of Context Cache Governance | Greyhound Standpoint | [quoted in The Deeper Read] |
| Greyhound Research (2025d)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Capacity And QuotaQuoted in The Deeper Read of Capacity And Quota | Greyhound Standpoint | [quoted in The Deeper Read] |
| Greyhound Standpoint | “could face underutilised infrastructure” (quoted in the text of The Deeper Read). | |
| Greyhound Research (2026a)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Agent Grid And Identity | Note | Tests for any agentic application, including naming each agent's identity and bounding its authority (summarized, not quoted). |
| Greyhound Research (2026b)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in What Ignite Changed, Lock-In And Exit | Note | Consolidation on one suite remains the right strategy for most enterprises, while the standard terms around it fall short (summarized, not quoted). |
| Schuman, E. (2026)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Lock-In And Exit. A Greyhound Research comment as publishedQuoted in The Deeper Read of Lock-In And Exit | Greyhound Research comment | [quoted in The Deeper Read] |
| Swain, G. (2026a)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Agents As Insiders. A Greyhound Research comment as published | Greyhound Research comment | “A sandbox can protect the host while failing to contain the work.” (quoted in the text of The Deeper Read). |
| Swain, G. (2026b)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Foundry On Customer AKS. A Greyhound Research comment as published | Greyhound Research comment | Self-hosting described as a staffing decision as much as an infrastructure one (quoted in the text of The Deeper Read). |
| Thomas, P. A. (2025)Read again November 19, 2026, between 00:10 and 00:45 UTCUsed in Capacity And Quota. A Greyhound Research comment as publishedQuoted in The Deeper Read of Capacity And Quota | Greyhound Research comment | [quoted in The Deeper Read] |
Part 9Appendix, Sources, And Disclosures
Sources And References: Every Document Behind Every Claim
Every fact, figure, and quotation in this research note traces to one of the sources below, presented in Harvard reference format and grouped by class. Each source is also linked in the sentence that carries it. Microsoft’s, AWS’s, and Google’s documents are each company’s account of its own products, prices, and terms, and prove what the company said rather than that the thing is true. The security research and news reports record what their authors found and were told, and the published comments listed under the research this note builds on are cited only for Greyhound Research’s own words as printed. Standing pages that carry no publication date are marked n.d. All sources were read between November 17 and November 19, 2026.
Microsoft announcements and blogs
Bhatia, V. (2026) What’s New in Azure Confidential Computing — Ignite 2026. Microsoft Community Hub, November 17. Available at: https://techcommunity.microsoft.com/blog/azureconfidentialcomputingblog/what%E2%80%99s-new-in-azure-confidential-computing-%E2%80%94-ignite-2026/4573909
Borkar, R. and Kishan, A. (2026) Azure Cobalt 200 VMs now available, proven on real workloads and built for the agentic era. Microsoft Azure Blog, November 17. Available at: https://azure.microsoft.com/en-us/blog/azure-cobalt-200-vms-now-available-proven-on-real-workloads-and-built-for-the-agentic-era/
Gallot, H. (2026) Securing the agentic insider. The Official Microsoft Blog, October 19. Available at: https://blogs.microsoft.com/blog/2026/10/19/securing-the-agentic-insider/
Hollan, J. et al. (2026) Introducing the Microsoft Copilot harness in Foundry Agent Service. Microsoft Foundry Blog, October 28. Available at: https://devblogs.microsoft.com/foundry/microsoft-copilot-harness-in-foundry/
Holeček, A. (2026) Security for the agentic era, powered by Microsoft Perception. Microsoft Security Blog, November 17. Available at: https://www.microsoft.com/en-us/security/blog/2026/11/17/security-for-the-agentic-era-powered-by-microsoft-perception/
Hulme, M. (2026) Designing agent-first platforms: What changes when agents do the work. Microsoft Azure Blog, September 23. Available at: https://azure.microsoft.com/en-us/blog/designing-agent-first-platforms-what-changes-when-agents-do-the-work/
Microsoft (2026a) Microsoft Ignite 2026 | Book of News. Microsoft News, November 17. Available at: https://news.microsoft.com/events/ignite/2026/book-of-news/
Microsoft (2026b) ACI Sandboxes: A Glimpse into the Future of Agentic Compute. Microsoft Community Hub, modified June 8. Available at: https://techcommunity.microsoft.com/blog/AzureCompute/aci-sandboxes-a-glimpse-into-the-future-of-agentic-compute/4526576
Microsoft (2026i) Preview of multiparty analytics with Azure Confidential Clean Rooms. Microsoft Community Hub, June 2. Available at: https://techcommunity.microsoft.com/blog/azureconfidentialcomputingblog/preview-of-multiparty-analytics-with-azure-confidential-clean-rooms/4523274
Microsoft (2026j) Run models and agents on your own Kubernetes clusters: AKS at Microsoft Ignite 2026. Microsoft Community Hub, November 17. Available at: https://techcommunity.microsoft.com/blog/appsonazureblog/run-models-and-agents-on-your-own-kubernetes-clusters-aks-at-microsoft-ignite-20/4573912
Nadella, S. (2026) Ignite 2026: Building the frontier ecosystem. The Official Microsoft Blog, November 17. Available at: https://blogs.microsoft.com/blog/2026/11/17/ignite-2026-building-the-frontier-ecosystem/
Nair, H. (2026) Introducing new experiences in Kubernetes Center at Microsoft Ignite 2026. Microsoft Community Hub, November 16. Available at: https://techcommunity.microsoft.com/blog/azureinfrastructureblog/introducing-new-experiences-in-kubernetes-center-at-microsoft-ignite-2026/4569688
Ulag, A. (2026) Microsoft Ignite 2026: What’s new in Azure, the platform purpose-built for AI. Microsoft Azure Blog, November 17. Available at: https://azure.microsoft.com/en-us/blog/microsoft-ignite-2026-whats-new-in-azure-the-platform-purpose-built-for-ai/
Winter, J. (2026) Introducing Azure Catalyst agentic modernization and new AI infrastructure innovations. Microsoft Azure Blog, November 17. Available at: https://azure.microsoft.com/en-us/blog/introducing-azure-catalyst-agentic-modernization-and-new-ai-infrastructure-innovations/
Microsoft documentation, repositories and sessions
Microsoft (2026c) Azure Container Apps Sandboxes overview. Microsoft Learn, updated November 6. Available at: https://learn.microsoft.com/en-us/azure/container-apps/sandboxes-overview
Microsoft (2026d) Egress policies and network controls for Azure Container Apps Sandboxes. Microsoft Learn, published May 21. Available at: https://learn.microsoft.com/en-us/azure/container-apps/sandboxes-egress-policies
Microsoft (2026e) Hosted agents in Foundry Agent Service. Microsoft Learn, updated November 11. Available at: https://learn.microsoft.com/en-us/azure/foundry/agents/concepts/hosted-agents
Microsoft (2026f) Prompt caching with Azure OpenAI in Microsoft Foundry Models. Microsoft Learn, updated November 6. Available at: https://learn.microsoft.com/en-us/azure/ai-foundry/openai/how-to/prompt-caching
Microsoft (2026g) Deploy and run Microsoft Foundry models and agents on AKS (BRK420). Microsoft Ignite 2026 session catalog, November 17. Available at: https://ignite.microsoft.com/en-US/sessions/1787856637165001EL2e-1791588324684001CZEV
Microsoft (2026h) New Foundry subscription shows 0 TPM quota for all Azure OpenAI models in every region - cannot deploy gpt-4o-mini. Microsoft Q&A, June 13. Available at: https://learn.microsoft.com/en-us/answers/questions/5919307/new-foundry-subscription-shows-0-tpm-quota-for-all
Microsoft (n.d.a) Getting Started with Azure Context Cache. GitHub, Azure/AzureContextCache. Available at: https://github.com/Azure/AzureContextCache
Microsoft (n.d.b) az storage context-cache container. Microsoft Learn, Azure CLI reference. Available at: https://learn.microsoft.com/en-us/cli/azure/storage/context-cache/container?view=azure-cli-latest
Microsoft (n.d.c) {AKS AI Manager} Add AKS AI Manager 2026-09-02-preview API (BYO cluster and custom models). GitHub, Azure/azure-rest-api-specs, pull request 46030. Available at: https://github.com/Azure/azure-rest-api-specs/pull/46030
Microsoft Security Response Center (2025) M365 Copilot Information Disclosure Vulnerability, CVE-2025-32711. Microsoft Security Update Guide, June 11. Available at: https://msrc.microsoft.com/update-guide/vulnerability/CVE-2025-32711
AWS
Amazon Web Services (2025) Amazon Bedrock AgentCore is now generally available. AWS What’s New, October 13. Available at: https://aws.amazon.com/about-aws/whats-new/2025/10/amazon-bedrock-agentcore-available
Amazon Web Services (2026a) Policy in Amazon Bedrock AgentCore is now generally available. AWS What’s New, March 3. Available at: https://aws.amazon.com/about-aws/whats-new/2026/03/policy-amazon-bedrock-agentcore-generally-available
Amazon Web Services (2026b) AWS re:Invent 2026. AWS Events. Available at: https://aws.amazon.com/events/reinvent/
Amazon Web Services (2026c) Nitro Sandboxes: Hardware isolation for every EKS pod, ECS task, and EC2 workload. AWS Compute Blog, November 10. Available at: https://aws.amazon.com/blogs/compute/nitro-sandboxes-hardware-isolation-for-every-eks-pod-ecs-task-and-ec2-workload/
Amazon Web Services (n.d.) Overview - Amazon Bedrock AgentCore. AWS Documentation. Available at: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html
Parthasarathy, M. (2026) Run agents securely and efficiently with the next generation AgentCore runtime. AWS Machine Learning Blog, November 12. Available at: https://aws.amazon.com/blogs/machine-learning/run-agents-securely-and-efficiently-with-the-next-generation-agentcore-runtime/
Google Cloud (2026a) Agent Identity overview. Google Cloud Documentation, updated November 17. Available at: https://docs.cloud.google.com/iam/docs/agent-identity-overview
Google Cloud (2026b) About Agent Sandbox on GKE. Google Cloud Documentation, updated November 17. Available at: https://docs.cloud.google.com/kubernetes-engine/docs/concepts/machine-learning/agent-sandbox
Google Cloud (2026c) Verifiable, private AI: Google Cloud expands Confidential Computing frontiers. Google Cloud Blog, June 23. Available at: https://cloud.google.com/blog/products/identity-security/verifiable-trust-in-the-ai-era-whats-new-in-confidential-computing
Google Cloud (2026d) Agent Gateway overview. Google Cloud Documentation, updated November 18. Available at: https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/gateways/agent-gateway-overview
Royal, B. and Hockin, T. (2026) Agent Sandbox on GKE is now available for everyone, and a first look at Agent Substrate. Google Cloud Blog, May 20. Available at: https://cloud.google.com/blog/products/containers-kubernetes/bringing-you-agent-sandbox-on-gke-and-agent-substrate
Model providers, chipmakers and open-source projects
Agentic AI Foundation (2026) A2A joins AAIF’s open agentic stack. AAIF Blog, August 17. Available at: https://aaif.io/blog/a2a-joins-aaif
Anthropic (n.d.) Prompt caching. Claude Platform Docs. Available at: https://platform.claude.com/docs/en/build-with-claude/prompt-caching
Cloud Native Computing Foundation (2026) Welcome llm-d to the CNCF: Evolving Kubernetes into SOTA AI infrastructure. CNCF Blog, March 24. Available at: https://www.cncf.io/blog/2026/03/24/welcome-llm-d-to-the-cncf-evolving-kubernetes-into-sota-ai-infrastructure/
Cloud Native Computing Foundation (n.d.) KAITO. CNCF Projects. Available at: https://www.cncf.io/projects/kaito/
Cloudflare (n.d.) Sandboxes. Cloudflare product page. Available at: https://www.cloudflare.com/products/sandboxes/
NVIDIA (2026) NVIDIA Kicks Off the Next Generation of AI With Rubin — Six New Chips, One Incredible AI Supercomputer. NVIDIA Newsroom, January 5. Available at: https://nvidianews.nvidia.com/news/rubin-platform-ai-supercomputer
OpenAI (n.d.) Prompt caching. OpenAI API documentation. Available at: https://developers.openai.com/api/docs/guides/prompt-caching
vLLM (n.d.) Automatic Prefix Caching. vLLM documentation. Available at: https://docs.vllm.ai/en/latest/features/automatic_prefix_caching/
Security research, standards and media
Amure, T. O. (2026) Microsoft CEO wants AI treated like a risky insider at work. TheStreet, October 11. Available at: https://www.thestreet.com/technology/microsoft-ceo-wants-ai-treated-like-a-risky-insider-at-work
Chismon, D. (2025) Prompt injection is not SQL injection (it may be worse). National Cyber Security Centre blog, December 8. Available at: https://www.ncsc.gov.uk/blog-post/prompt-injection-is-not-sql-injection
Help Net Security (2026) OWASP 2026 LLM Top 10: ‘The model will be fooled’. Help Net Security, August 6. Available at: https://www.helpnetsecurity.com/2026/08/06/owasp-2026-llm-top-10-released/
Infosecurity Magazine (n.d.) UK NCSC Raises Alarms Over Prompt Injection Attacks. Available at: https://www.infosecurity-magazine.com/news/ncsc-raises-alarms-prompt/
Lakshmanan, R. (2025) Zero-Click AI Vulnerability Exposes Microsoft 365 Copilot Data Without User Interaction. The Hacker News, June 12. Available at: https://thehackernews.com/2025/06/zero-click-ai-vulnerability-exposes.html
Riley, D. (2026) Microsoft targets the ‘agentic insider’ with new controls for rogue AI agents. SiliconANGLE, October 19. Available at: https://siliconangle.com/2026/10/19/microsoft-targets-the-agentic-insider-with-new-controls-for-rogue-ai-agents/
TEE.fail researchers (n.d.) TEE.fail: Breaking Trusted Execution Environments via DDR5 Memory Bus Interposition. Research site. Available at: https://tee.fail/
The New Stack (2026) AWS, Google Cloud, Microsoft Azure, and Cloudflare now all offer agent sandboxes. None built them the same way. The New Stack, July 24. Available at: https://thenewstack.io/cloud-agent-code-sandboxes/
Toulas, B. (2025) TEE.Fail attack breaks confidential computing on Intel, AMD, NVIDIA CPUs. BleepingComputer, October 28. Available at: https://www.bleepingcomputer.com/news/security/teefail-attack-breaks-confidential-computing-on-intel-amd-nvidia-cpus/
Research this note builds on
Greyhound Research (2025a) Microsoft Build 2025: Making AI Agents More Capable and Secure. Research note, May 22. Available at: https://greyhoundresearch.com/microsoft-build-2025-making-ai-agents-more-capable-and-secure/
Greyhound Research (2025b) Microsoft Build 2025: Supporting the Open Agentic Web. Research note, May 26. Available at: https://greyhoundresearch.com/microsoft-build-2025-supporting-the-open-agentic-web/
Greyhound Research (2025c) Transforming AI Economics with Procedural Memory. Research note, August 30. Available at: https://greyhoundresearch.com/transforming-ai-economics-with-procedural-memory/
Greyhound Research (2025d) Microsoft Layoffs: AI’s Impact on Sales Dynamics. Research note, June 23. Available at: https://greyhoundresearch.com/microsoft-layoffs-ais-impact-on-sales-dynamics/
Greyhound Research (2026a) Everyone Is Building AI Agents. Oracle Is Moving The Contest Inside The Runtime. Research report, July 20. Available at: https://greyhoundresearch.com/everyone-is-building-ai-agents-oracle-is-moving-the-contest-inside-the-runtime/
Greyhound Research (2026b) Microsoft 365 Outage – When The Suite Goes Dark, Restoration Order Becomes The Control. Research note, November 15. Available at: https://greyhoundresearch.com/microsoft-365-outage-when-the-suite-goes-dark-restoration-order-becomes-the-control/
Schuman, E. (2026) Microsoft, OpenAI change contract terms — again. Computerworld, April 27. Cited for a Greyhound Research comment as published. Available at: https://www.computerworld.com/article/4163971/microsoft-openai-change-contract-terms-again.html
Swain, G. (2026a) OpenAI agent swarm exposes a blind spot in AI containment. CIO, September 8. Cited for a Greyhound Research comment as published. Available at: https://www.cio.com/article/4219643/openai-agent-swarm-exposes-a-blind-spot-in-ai-containment.html
Swain, G. (2026b) Microsoft, Nvidia push provenance checks as open-weight AI adoption accelerates. CSO Online, October 19. Cited for a Greyhound Research comment as published. Available at: https://www.csoonline.com/article/4237095/microsoft-nvidia-push-provenance-checks-as-open-weight-ai-adoption-accelerates.html
Thomas, P. A. (2025) Microsoft loses two senior AI infrastructure leaders as data center pressures mount. Computerworld, November 26. Cited for a Greyhound Research comment as published. Available at: https://www.computerworld.com/article/4096759/microsoft-loses-two-senior-ai-infrastructure-leaders-as-data-center-pressures-mount-2.html
Part 9Appendix, Sources, And Disclosures
Disclosures: Independence, Financial, Intended Use, Methodology, Authoring, And Copyright
The evidence in this research note is current through November 19, 2026, with an evidence cutoff of 00:45 UTC that day, when the sources were checked again for this edition. The first edition’s cutoff was 02:00 UTC on November 18, the day after Microsoft’s Ignite announcements. At both times, searches of Microsoft Learn found no documentation for Agent Grid or the Agent Sandbox extension for AKS. Any written response received from an organization named in it after that time is dated where it appears.
Independence Disclosure
Research under the Greyhound Research brand is independent, and no organization holds a right of approval over its content or publication. It is not commissioned, sponsored, or funded by any organization it covers. Greyhound Research may send a research note to companies named in it for factual review before publication. Any such review is confined to the correction of fact, and no viewpoint, conclusion, or recommendation is guided by those companies. All insights and recommendations are Greyhound Research’s alone, and editorial control remains entirely with Greyhound Research. Greyhound Consilium is the B2B thought-leadership arm of The House Of Greyhound and a sister brand to Greyhound Research. It builds on Greyhound Research’s independent research through collaborations that produce thought-leadership content, research, and practical toolkits for decision-makers. Every Greyhound Consilium publication names its partner on the face of the document. This research note is published under the Greyhound Research brand.
Financial Disclosure
Greyhound Research undertakes advisory and consulting engagements with enterprise buyers and technology providers, and no such engagement confers control over its research findings or its decision to publish them. For this research note, Greyhound Research has no financial interest in, and no commercial relationship with, Microsoft, Amazon Web Services, Google, Anthropic, or OpenAI to declare. Oracle is a client of Greyhound Research. Oracle is not a subject of this research note.
Intended Use
This research note is written for chief information officers, chief information security officers and identity leads, heads of platform engineering and enterprise architects, FinOps leads and chief financial officers, data protection officers, procurement and vendor management leaders, and the press. It reads Microsoft’s announcements at Microsoft Ignite 2026, the documentation beside them, and the published documentation of AWS, Google, model providers, and open-source projects, to inform decisions on agent runtime, identity, policy, caching, placement, deployment, responsibility, capacity, dependency, and exit. This research note is not investment advice, a recommendation to buy, sell, or hold any security, or a view on any company’s share price or valuation. It is not legal advice, and its reading of contract and data protection matters is to be confirmed with counsel. Its assessments rest on the documents listed in Sources And References as they stood on November 19, 2026, and can change as the companies named publish more. Product availability, status, regions, features, and prices are described as each company published them through November 19, 2026, so confirm them with the vendor before acting on them.
Research Methodology
Greyhound Research’s analysis combines lived experience with public sources. The lived experience is years of continuously tracking the market, Greyhound Fieldnotes from ongoing advisory engagements with enterprise buyers and with technology providers, and Greyhound Pulse, the firm’s market trackers. Lived experience forms the judgment and never stands in for evidence. The public sources are the sixty-two documents listed in Sources And References, in seven classes, and every factual claim is verified against the document that supports it. The passages relied on are quoted in The Deeper Read of the section that uses them, and the other words taken from each source, the passages summarized rather than quoted and the reading notes, are set out in the appendix. A finding that something is not stated, not documented, or not found is bounded to the documents and searches the text names, as they stood at 00:45 UTC on November 19, 2026. Statuses are given as each company states them, and where two documents from one company word the same fact differently, both readings are given.
Start-up and allocation figures from Microsoft, AWS, and Google are each company’s own claims, measured on different bases, and are not presented as directly comparable. Cache prices are the providers’ published rates, and the shares of base input shown in Exhibit 5 restate each provider’s own multiplier. No composite figure is computed. The Greyhound CIO Pulse 2025 figures are cited as published in Greyhound Research’s research notes of May 22 and May 26, 2025, labeled with their edition. No Greyhound Fieldnotes are quoted, and no factual claim in this research note rests on unpublished material.
The three-route reading of agent harm, the readings in Exhibits 1 to 11 and Tables 1 to 6, the risks, the Greyhound CXO Runbook, the questions for Microsoft, the signals, and every term proposed for a contract are Greyhound Research’s analysis, not statements by Microsoft or any other company. Quotations are checked against the live source before publication.
Authoring Disclosure
This material reflects the analysis, viewpoint and editorial judgment of Greyhound Research. The position presented is developed from Greyhound Research’s research, advisory work and accumulated institutional knowledge. Depending on the assignment, AI may be used to retrieve relevant material; structure, draft or polish prose from analyst-provided concepts, dictation, notes, rough drafts and source material; support secondary research and factual cross-checking; and assist with editing, condensation and consistency with Greyhound Research’s house style. AI does not independently establish Greyhound Research’s viewpoint, analytical position or conclusions. Material factual claims are verified against source evidence before release. Greyhound Research reviews and approves the final material and retains responsibility for it.
Copyright Policy
All content in this document and on the Greyhound Research website is protected by copyright law and may not be reproduced, distributed, transmitted, displayed, published, or broadcast without the prior written permission of Greyhound Research or, in the case of third-party materials, the prior written consent of the copyright owner of that content. You may not alter, delete, obscure, or conceal any trademark, copyright, or other notice appearing in any Greyhound Research content. Quotation of up to one hundred words is permitted where it carries attribution to Greyhound Research and a link to the published page. Reproduction, redistribution or republication beyond that extent, in whole or in part, by email or in any medium including websites, newsletters and intranets, requires prior written permission, which may be requested at the address below. The published version of each content piece on the Greyhound Research website carries sharing tools for readers who wish to share it.
For questions arising from these disclosures, including licensing requests, contact the Greyhound Research Community Relations Team at connect@thofgr.com. For media inquiries, write to media@thofgr.com.

Analyst In Focus: Sanchit Vir Gogia
Sanchit Vir Gogia, or SVG as he is popularly known, is a globally recognized technology analyst, innovation strategist, digital consultant and board advisor. SVG is the Chief Analyst at Greyhound Research (part of The House Of Greyhound), a Global, Award-Winning Technology Research, Advisory, Consulting & Education firm. Greyhound Research works closely with global organizations, their CxOs and the Board of Directors on Technology & Digital Transformation decisions.
Discover more from Greyhound Research
Subscribe to get the latest posts sent to your email.
