Microsoft Agent Identity – When Rivals’ Agents Get A Microsoft Name, Enforcement Becomes The Control

Publication Record

Evidence current through: Thursday, November 19, 2026, at 12:40 a.m. UTC (7:40 p.m. ET and 4:40 p.m. PT on Wednesday, November 18).

Most of the agent identity and governance items Microsoft announced at Ignite on November 17, 2026, are in preview, and the note reads each at the status Microsoft gave it by the time above. Where two Microsoft pages disagree, both are reported. Developments after that time 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 set out to become the one place where a company keeps track of all its AI agents, whoever built them. A feature called connected platforms copies agents from Salesforce, AWS, Google, Anthropic, and other vendors into Agent 365, Microsoft’s agent management product, and gives each one an Entra Agent ID, an entry in Microsoft’s identity directory. Agents running on employees’ own computers, such as Claude Code and GitHub Copilot CLI, can get an ID too. Most of it is in preview, a test release without full service terms.

Greyhound Research’s position is that Microsoft has built the most complete directory for agents it does not run, and that a directory entry is not a control. Microsoft’s own documentation says the link to rivals’ platforms is read-only. Agent 365 can name and watch a Salesforce or AWS agent. Its documentation gives it no way to stop one. When rivals’ agents get a Microsoft name, enforcement becomes the control, and enforcement stays wherever the agent’s own credentials are issued. Greyhound Research reads this as an early and familiar stage. Identity products have often listed systems before they could govern them, and buyers who treat the list as a starting point rather than as cover are likely to fare best.

Three findings carry the argument. First, the “50 million agents” Microsoft cites has three wordings and no published definition, so it measures registrations, not governed agents. Second, Microsoft names eleven connected platforms, while its documentation lists seven, syncs only when an administrator asks, and does not promise to remove agents deleted at the source. Third, on employees’ computers Microsoft does enforce, through Defender, Intune, and Windows containers, but its own page warns that a block can arrive after a tool has already run.

The note sets out a ten-move Greyhound CXO Runbook, with seven exhibits, seven tables, and six questions for Microsoft. None of the moves waits on Microsoft. The first three are simple to state. Register every agent as a privileged account with a named sponsor. Map where each agent’s credentials are issued, because that is where it can be stopped. Keep tool-level limits on each platform or in a gateway the company controls.

The evidence has limits. Most of the items are previews, and Microsoft’s announcement and its documentation disagree in places, which this note records. Greyhound Research names the signals that would change its view. Until they arrive, the advice holds: use Agent 365 as the inventory of record, and keep the power to stop each agent with the platform that issues its credentials.

In This Research NoteSection guide · AI drafted · Analyst reviewed

Seventeen sections in eight 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 twelve 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.

  1. Part 1What Ignite Changed
  2. 1.1What Microsoft Announced: One Directory For Every Vendor’s Agents, Mostly In Preview

    Connected platforms, Agent IDs for local agents, and endpoint controls. Two pieces are generally available. Most of the rest is in preview, and the connected platform rollout completes in December.

  3. 1.2The 50 Million: A Count Without A Definition

    Microsoft’s headline number appears in three wordings across three weeks. None says whether it counts active agents, unique agents, or agents from other vendors.

  4. Part 2The Plane And Its Limits
  5. 2.1A Name Is Not A Control: What Connected Platforms Actually Do

    For rivals’ platforms, Agent 365 copies the list and reads the activity. Microsoft’s documentation says it does not act on the source platform and cannot add controls the platform lacks.

  6. 2.2The Connected Roster: Eleven Platforms Named, Seven Documented

    The announcement names eleven platforms. The documentation lists seven, syncs only on request, needs a stored key for each, and does not promise to remove agents deleted at the source.

  7. 2.3Where Enforcement Lives: One Agent, Two Rulebooks

    An agent can be stopped wherever its credentials are issued. Agents that sign in through Entra can be stopped by Entra. Agents that act under their own platform’s credentials can be stopped only there.

  8. Part 3Local Agents And Liability
  9. 3.1Local And Shadow Agents: The Laptop Becomes The Checkpoint

    On managed Windows devices Microsoft can find, name, contain, and inspect agents such as Claude Code and Copilot CLI. Discovery and runtime protection are released. Naming and blocking are in preview.

  10. 3.2When Enforcement Fails: Who Answers For A Rival’s Agent

    Microsoft’s documents keep it out of enforcement on rivals’ platforms. If a listed agent misbehaves, the platform that issued its credentials and the customer who set its rights carry the consequences.

  11. Part 4The Platform Contest
  12. 4.1Every Vendor Wants The Directory: A Year Of Rival Bids

    In 2026 AWS, Google, Okta, and ServiceNow each offered to register or govern agents from other platforms. Microsoft’s bid is the broadest by named platforms and the only one tied to the workforce directory most companies already run.

  13. 4.2Three Clouds Compared: Microsoft Leads On Reach, AWS On Enforcement

    Microsoft, AWS, and Google each give agents an identity. Microsoft reaches furthest across rivals and devices. AWS has generally available rules on each tool call. Google documents such rules without a single release label and uses an open identity standard.

  14. Part 5The Risks And The Counter-Argument
  15. 5.1Risks To Watch: Where A Cross-Vendor Plane Can Fail

    Five risks, from a name mistaken for a control to liability no contract assigns, each with an early warning and an owner.

  16. 5.2The Strongest Objection: One List Is Better Than None

    The fair counter-case: inventory is the first control, Microsoft holds the workforce directory, and observation beats blindness. It holds for inventory, and it weakens if the list is read as enforcement.

  17. Part 6What To Do, By Role
  18. 6.1The Greyhound CXO Runbook: Ten Moves, None Waiting For Microsoft

    Ten moves for the CISO, identity leads, enterprise architects, endpoint leads, general counsel and procurement, the CIO, and the audit committee, each naming the assumption it corrects and who decides.

  19. 6.2Questions For Microsoft: What Buyers Will Ask Next

    Six questions only Microsoft can answer, on the 50 million, enforcement per platform, sign-in, the roster, runtime protection, and responsibility. Each is tied to a Runbook move.

  20. Part 7The Final Greyhound Standpoint
  21. 7.1Agent Identity: The Final Greyhound Standpoint

    Microsoft has the broadest agent directory, its enforcement stops at its own products and devices, and the pass is the control. The limits of the evidence, and the signals that would change the view within twelve months.

  22. Part 8Appendix, Sources, And Disclosures
  23. 8.1Appendix: The Evidence Register, Source By Source

    Every source behind this research note, with when it was read and where it is used. Passages quoted in The Deeper Read are not repeated here.

  24. 8.2Sources And References: Every Document Behind Every Claim

    Thirty-three sources in six classes, from Microsoft’s announcements and documentation to AWS, Google, Okta, ServiceNow, Zenity, the press, and the research this note builds on.

  25. 8.3Disclosures: 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

Part 1What Ignite Changed

What Microsoft Announced: One Directory For Every Vendor’s Agents, Mostly In Preview

Greyhound SpotlightMicrosoft’s announcement says connected platforms use secretless sign-in. Its own documentation, last updated November 6, says every supported platform still needs a stored key.

Microsoft’s announcement for Agent 365 at Ignite makes a large offer. Connect a rival’s agent platform once, and its agents appear in Microsoft’s registry with an Entra Agent ID. Microsoft says the rollout covers Salesforce Agentforce, Amazon AgentCore, Google Vertex AI, Databricks, Snowflake, Anthropic, OpenAI, Oracle, SAP, ServiceNow, and UiPath, and will complete in December. For an identity team, that is a bid to make Microsoft’s directory the list of every agent the company runs.

For identity teams, this is a change of scope more than a change of product. Directories have for decades listed the people and systems a company runs itself, while each agent platform has kept its own list of agents. Ignite asks Microsoft’s directory to hold its rivals’ lists as well, in preview now and with the rollout due to complete in December. Greyhound Research expects the first decision to arrive quietly, as a connector one administrator switches on, rather than as a program a steering group approves. What tends to be missed is that each connection is also a new standing credential into a rival’s platform, and someone in the company has to own it.

Table 1 sets the Ignite items beside the products they build on, with the status Microsoft gave each one. Generally available means fully released, with support and service terms.

Greyhound Research · Table 1 · Status ledger

Most of the cross-vendor pieces are previews

Microsoft’s Ignite 2026 agent identity and governance items, with the products they build on, at the status Microsoft gave each, November 19, 2026.

Greyhound Research’s judgment: pilot the cross-vendor inventory now, and claim no cross-vendor control until each platform’s documentation supports it.

Announced at Ignite, in previewGenerally availableBase product
OfferingStatusWhat Microsoft has writtenNot yet written
Connected platforms in Agent 365PreviewAgents on rival platforms synced into the registry with an Entra Agent IDRollout completes in December, capabilities vary by platformWhat is enforced on each platform
Agent ID for discovered local agentsPreview“to separate agent from human activity”Starting with GitHub Copilot CLI and Claude CodeWhich local agents follow
Agent ID for Autopilot agentsPreviewMicrosoft’s own personal agent gets an Agent IDRelease date
Intune local AI agent governancePreviewDevices with unapproved agents marked noncompliant and blocked from company dataRelease date
Global Secure Access for local agents, MCP Firewall, unsanctioned agentsPreviewNetwork control over where agents go and which tools they reachRelease date
Dynamic Agent Risk and dynamic permissionsPreviewRisk scores for each agentTask-scoped access in limited previewRelease date
Action centerPreview in DecemberSurfaces inactive and ownerless agentsPrivate preview, per the Book of NewsWhich pages are right
Local agent discovery in DefenderGenerally availableMore than 60 local AI agents found on managed devicesCoverage of unmanaged devices
Runtime protection for local agentsGenerally available on WindowsInspects the agent loop, in audit or block modeIts Microsoft Learn page is titled previewA block before every tool runs
Agentic traffic acquisitionGenerally availableRoutes agent traffic for visibility and enforcementWhich agents it covers
Microsoft Execution ContainersGenerally availableA contained workspace per agent on Windows 11Since October 2026Coverage beyond Windows
Agent 365Generally availableSold since May 1, 2026, at USD 15 per userIncluded in Microsoft 365 E7No price change found at Ignite
Microsoft Entra Agent IDAvailableAvailable to all Entra customersSecurity features require Agent 365A single release label
Against this: Agent 365 has been generally available since May 1, 2026, so the new pieces extend a product already in use.
Source: Microsoft Book of News (November 17, 2026), Microsoft Tech Community (Agent 365, November 17, 2026), Microsoft Security Blog and Windows Experience Blog (November 17, 2026), Official Microsoft Blog (March 9, 2026), Microsoft Learn.

Table 1 lists Microsoft’s agent identity and governance items by status, and shows most cross-vendor items in preview. In preview: connected platforms in Agent 365, Agent ID for discovered local agents starting with GitHub Copilot CLI and Claude Code, Agent ID for Autopilot agents, Intune local AI agent governance, Global Secure Access for local agents with the MCP Firewall, Dynamic Agent Risk with dynamic permissions in limited preview, and the action center in December. Generally available: local agent discovery in Defender, runtime protection for local agents on Windows, whose Learn page is titled preview, agentic traffic acquisition, and Microsoft Execution Containers. Base products: Agent 365, sold since May 1, 2026, at USD 15 per user, and Entra Agent ID, available to all Entra customers. Statuses are as Microsoft published them through November 19, 2026.

Table 1. Most of the cross-vendor pieces are previews. Data as of November 19, 2026. © Greyhound Research.

Two of the new items are generally available: discovery of local agents by Defender, and Defender’s protection of those agents while they run on Windows. Microsoft Execution Containers have been generally available since October. Almost everything that reaches beyond Microsoft’s own products is in preview, and three items carry different statuses on different Microsoft pages. Agent 365 itself has been sold since May 1 at USD 15 per user, so the base product is already in use.

Greyhound Standpoint:

At Greyhound Research, we believe Ignite turned Agent 365 from a register of Microsoft’s own agents into a bid to hold the register of every vendor’s agents, and that the ambition is the right one even though the documents have not yet caught up with it. Companies already run agents on several platforms, and none of those platforms lists the agents of the others. One list would give security teams a starting point most of them lack today.

The difficulty is the order of release. The pieces that reach beyond Microsoft’s own products are previews, and the announcement and the documentation describe the connection differently. Identity products have moved this way before. Directory and single sign-on suites tended to launch with a long list of supported systems, and the depth of each connector became clear only once customers ran it. The gap usually narrowed, at a pace set by the documentation rather than the keynote. Microsoft deserves credit for shipping the base product first. Agent 365 has been sold since May, and two of the new local agent controls are generally available. A per-platform statement of what is enforced, published before the December rollout completes, would strengthen this view.

Identity leads should run connected platforms as an inventory pilot this quarter on two or three platforms, and the CISO should hold any board statement that rivals’ agents are governed until that statement exists.

The Deeper Read · About 3 minutes

Select a question to open the full answer

1What exactly did Microsoft announce for connected platforms?

In Microsoft’s words. The Agent 365 post of November 17 says agents from Google, Salesforce, Databricks, Snowflake, Anthropic, OpenAI, and more “now get their own identity in Microsoft Entra,” marked preview with capabilities that vary by platform. It says connected agents automatically receive an Entra Agent ID, and that the rollout will complete in December.

How the connection signs in. The same post says secretless authentication uses workload identity federation, a way to sign in with a token from another system rather than a stored password. Microsoft’s connected platforms documentation, updated November 6, says the opposite for every platform it lists. Greyhound Research reads the documentation as the older of the two statements and records the conflict rather than resolving it.

What the two descriptions tell a buyer. A gap between an announcement and its documentation is common in a preview, and it is useful rather than alarming, because it shows which document to work from. Engineers will configure what the documentation says, and today that is a stored key for every platform. Until the pages agree, buyers should plan, budget, and assess risk on the documented method, and treat keyless sign-in as a promised improvement to be confirmed in writing before a pilot is approved.

In the source’s own words

  • Microsoft Tech Community, Agent 365 at IgniteConnected agents now automatically receive a Microsoft Entra Agent ID, which brings them under identity managementSource
  • Microsoft Tech Community, Agent 365 at IgniteSecretless authentication uses Microsoft Entra workload identity federation to connect Agent 365 to supported platformsSource
  • Microsoft Learn, connected platformsEvery currently supported platform requires a key-based credential, such as an API key, secret, or private key.Source
2What else did Ignite add for identity and endpoints?

Entra and Intune. The Book of News says Agent ID for discovered local agents, in preview, will assign an Agent ID to local agents, starting with GitHub Copilot CLI and Claude Code. Intune’s local AI agent governance, also in preview, will mark devices running unapproved agents as noncompliant and use Conditional Access, Microsoft’s sign-in rule engine, to block them from company data. The MCP Firewall in Global Secure Access, in preview, will control access to the servers through which agents reach tools.

The base product. Judson Althoff, CEO of Microsoft’s commercial business, announced in March that Agent 365 would be generally available on May 1, 2026, at USD 15 per user, and called it the control plane for AI agents. Entra Agent ID itself is available to all Entra customers, while extending Entra’s security features to agents requires Agent 365.

In the source’s own words

  • Microsoft’s Book of News, Microsoft Entrastarting with GitHub Copilot CLI and Claude Code, to separate agent from human activity and help secure their access.Source
  • Microsoft’s Book of News, Microsoft Entrawill control access to MCP servers, tools, resources, prompts, protocols and transports.Source
  • The Official Microsoft Blog, Althoff on March 9General availability of Agent 365 on May 1 for $15 per userSource
  • Microsoft Learn, What is Microsoft Entra Agent IDExtending Microsoft Entra security features to agents requires Microsoft Agent 365.Source
3Where do Microsoft’s pages disagree on status?

Three items, two statuses each. Defender’s runtime protection for local agents is generally available on Windows in the Book of News and the Agent 365 post, while its Microsoft Learn page is titled preview and carries a prerelease notice. The action center is expected in private preview in December in the Book of News and in preview in December in the Agent 365 post. The sign-in method for connected platforms differs as described above. This note gives the status each page states.

In the source’s own words

  • Microsoft Learn, AI agent runtime protectionSome information in this article relates to a prereleased product which may be substantially modified before it's commercially released.Source
  • Microsoft’s Book of News, Agent 365 governanceThe action center is expected to enter private preview in December.Source

↑ Back to the section guide

Part 1What Ignite Changed

The 50 Million: A Count Without A Definition

Greyhound SpotlightMicrosoft’s “50 million agents” was already public three weeks before Ignite. On October 28 Satya Nadella put it as agents “now in” Agent 365, up 30% in the quarter. At Ignite it became agents “registered” since May.

Microsoft CEO Satya Nadella said in his Ignite keynote that Agent 365 already holds over 50 million agents. The number is meant to show that the directory is the market’s default. A count is only as useful as its definition, and Microsoft has not published one. Registered is not the same as active, and an agent built in a few clicks and not yet used counts the same as one that runs a finance process.

The figure matters less for what it says about Microsoft than for how it will be used inside buyers’ organizations. Vendor counts have long found their way into internal approvals and board papers as proof that a product is the market default, and a headline of 50 million is likely to travel the same way. What is new is the thing being counted. A seat count measured licensed people, each with a manager and a leaving date. An agent count can include items made in a few clicks, tested once, and left. The quieter consequence sits with the buyer: a company that cannot produce its own count of active, owned agents has no figure to set against Microsoft’s.

Table 2 sets out the three wordings of the figure and what each leaves open.

Greyhound Research · Table 2 · One figure, three wordings

The 50 million has no published definition

Each public wording of Microsoft’s count of agents in Agent 365, October 28 to November 17, 2026, and what none of them states.

Read the 50 million as a measure of how easily agents are made, not of how many are governed.

A published wordingNot published
DateWho said itThe wordingWhat it leaves open
October 28, 2026Nadella, after the earnings call“now in A365”Up 30% in the quarterRegistered or active
November 17, 2026Nadella’s Ignite keynote“over 50 million agents in Agent 365”Active, unique, or by source
November 17, 2026Microsoft’s Agent 365 post“registered” since the May 2026 launchActive, duplicates, or by vendor
Not publishedMicrosoftAgents active in the last 30 daysNot stated
Not publishedMicrosoftAgents by source platform, Microsoft or rivalNot stated
Not publishedMicrosoftAgents with a named owner or sponsorNot stated
Against this: Microsoft is building an action center to surface inactive and ownerless agents, which shows it knows the registry needs cleaning.
Source: Satya Nadella (October 28, 2026), Official Microsoft Blog (Nadella keynote remarks, November 17, 2026), Microsoft Tech Community (Agent 365, November 17, 2026).

Table 2 sets out three public wordings of Microsoft’s figure of 50 million agents, and shows that none defines it. On October 28, 2026, after Microsoft’s quarterly earnings call, Nadella wrote that over 50 million agents are now in Agent 365, up 30% in the quarter. On November 17, Nadella’s Ignite keynote, as Microsoft posted it, said over 50 million agents are in Agent 365. The same day, the Agent 365 post said more than 50 million were registered since the May 2026 launch. Microsoft has not published how many were active in the last 30 days, how many come from each source platform, or how many have a named owner. The figure is Microsoft’s claim.

Table 2. The 50 million has no published definition. Data as of November 19, 2026. © Greyhound Research.

The three wordings describe the same number differently. Nadella’s October wording gives a growth rate for the quarter, his Ignite wording gives no period, and the Agent 365 post counts agents registered since its May launch. None says how many are active, how many are duplicates, or how many come from other vendors’ platforms rather than Microsoft’s own tools. The figure is Microsoft’s claim and is not used in this note’s analysis.

Greyhound Standpoint:

Greyhound Research’s view is that the 50 million measures how easily agents are created in Microsoft’s tools, not how many agents are governed. A registration is created when an agent is built, and it stays until someone removes it. In a product where agents take minutes to make, the count will grow faster than any team’s ability to review what it counts.

The pattern is familiar from earlier waves of self-service technology. Collaboration sites, cloud subscriptions, and service accounts all multiplied once users could create them without a ticket, and each wave left a backlog of unowned objects that someone later had to find and retire. Microsoft appears to expect the same, since its action center begins by surfacing inactive and ownerless agents. The number deserves some credit, because a registry that captures agents as they are made is better than one that learns of them later. Its limit is the missing definition. Three wordings in three weeks do not say whether the count is active, unique, or drawn from other vendors’ platforms. A published split by active and registered agents, and by source, would make it a governance figure.

CIOs should keep the figure out of any business case. At the next quarterly review, they should report their own count of active agents with a named owner, on every platform, and ask Microsoft for the same split for their tenant.

The Deeper Read · About 2 minutes

Select a question to open the full answer

1Where does each wording come from?

Before the launch. In March, before general availability, Judson Althoff wrote that in just two months, tens of millions of agents had appeared in the Agent 365 Registry. Greyhound Research’s reading is that part of the 50 million may therefore predate the May launch from which the Agent 365 post counts. Microsoft has not said.

October 28. After Microsoft’s quarterly earnings call that day, Nadella wrote that over 50 million agents are now in Agent 365, up 30% quarter over quarter. He did not say whether the count is of registered or active agents.

November 17. In his Ignite keynote, which Microsoft posted as edited remarks, Nadella said Microsoft already has over 50 million agents in Agent 365, and called Agent 365 a control plane that extends existing governance, identity, security, and management systems. The Agent 365 post of the same day says organizations have registered more than 50 million agents since its May 2026 launch.

Why the drift in wording matters. Each version is accurate on its own terms. Read in sequence, they move from agents “now in” Agent 365, to agents in it with no period stated, to agents registered since a launch date. That drift suggests a number chosen to show momentum rather than to support management. A buyer cannot reconcile any of the three with its own records, so the useful response is to ask Microsoft for the split in its own tenant and to treat the headline as background.

In the source’s own words

  • The Official Microsoft Blog, Althoff on March 9In just two months, tens of millions of agents have appeared in the Agent 365 Registry.Source
  • Satya Nadella, October 28Over 50M agents are now in A365, up 30% QoQ.Source
  • The Official Microsoft Blog, Nadella at IgniteWe already have over 50 million agents in Agent 365.Source
  • Microsoft Tech Community, Agent 365 at IgniteOrganizations have registered more than 50 million agents in Microsoft Agent 365 since its May 2026 launch.Source
2What does Microsoft’s own tooling say about the registry’s quality?

Cleanup is on the roadmap. The Agent 365 post says the action center will begin by surfacing inactive and ownerless agents for review and cleanup, in preview in December. A registry that needs a cleanup tool is a registry with agents nobody is watching. Greyhound Research found in May 2025, in Microsoft Build 2025: Making AI Agents More Capable and Secure, that 69% of CIOs surveyed for Greyhound CIO Pulse 2025 named silent privilege creep as their top concern with AI agents. An ownerless agent is where that creep goes unnoticed.

One customer’s count. Microsoft’s October 29 customer story on Insight Enterprises says that as of September 2026 Insight had discovered 9,734 agents across its environment. Will Pocknell, Insight’s Director of Cybersecurity, is quoted there on what that scale demands. One company with close to ten thousand agents shows why a count needs a definition before it can support a governance claim.

In the source’s own words

  • Microsoft Tech Community, Agent 365 at IgniteIt begins by surfacing inactive and ownerless agents for review and cleanup.Source
  • Microsoft customer story, Insight EnterprisesManaging AI at scale requires more than seeing individual tools or agents in isolation.Source
  • Greyhound Research, Build 2025 agents noteAccording to Greyhound CIO Pulse 2025, 69% of CIOs surveyed now say their top concern with AI agents is not performance or hallucinationSource

↑ Back to the section guide

Part 2The Plane And Its Limits

A Name Is Not A Control: What Connected Platforms Actually Do

Greyhound SpotlightMicrosoft’s documentation says Agent 365 “never invokes, executes, or interacts with” agents on a connected platform. It can list a Salesforce agent. It is not designed to stop one.

Connected platforms sounds like central control. The documentation describes something narrower. An administrator gives Agent 365 a key to the rival platform. Agent 365 copies the list of agents into its registry and, where the platform allows, reads their activity records. It does not change anything on the other platform, and it cannot add controls the platform does not already have. The power to stop a Salesforce agent stays in Salesforce.

The design choice is sensible, and it says something about where the market stands. A read-only link is the least intrusive way for one vendor to connect to another’s platform, because it asks the rival for little more than a key and its logs. Anything stronger would need each rival to grant Microsoft powers on its platform, and no document read for this note describes such an arrangement. In Greyhound Research’s view, platform owners will keep the power to stop agents on their own ground for some time. For buyers, the practical effect is organizational. The team that spots a problem in Agent 365 and the team that can act on it in Salesforce may sit in different departments.

Exhibit 1 sets six controls against three kinds of agent: those Microsoft builds and runs, those on a connected platform, and local agents on employees’ computers. A control is enforced when Microsoft’s products can stop the agent, and observed when they can only see it.

Greyhound Research · Exhibit 1 · Seeing against stopping

On rivals’ platforms, Microsoft can see but not stop

Five controls against three kinds of agent, marked by what Microsoft’s own documents say its products can do, November 19, 2026.

Greyhound Research’s judgment: on rivals’ platforms Agent 365 is an inventory. Report those agents as observed, not governed.

Can act, documentedCan act, in previewCan only seeNot documentedMS Agents Microsoft runs · CP Connected platforms · LA Local agents
What Microsoft’s products can do
Agents Microsoft runs
Agents on connected platforms
Local agents on managed Windows devices
1List the agent in the registry
MSDocumented
CPPreview, read-only
LAPreview
2Read what the agent does
MSDocumented
CPFrontier preview, five of seven
LAReleased
3Give the agent an Entra Agent ID
MSDocumented
CPAnnounced, not in documentation
LAPreview
4Refuse the agent access when it signs in
MSDocumented
CPSees only
LAPreview, by device
5Confine or block the agent while it acts
MSNot assessed here
CPNot documented
LAReleased on Windows
Against this: An agent on a rival platform that is also wired to sign in through Entra can be refused access by Entra for systems that trust Entra.
Source: Greyhound Research analytical model, not measured data. Greyhound Research reading of Microsoft Learn (connected platforms, Entra Agent ID, runtime protection), the Book of News, and Microsoft’s Agent 365 post.

Exhibit 1 is a matrix of five controls against three kinds of agent, and shows that on connected platforms Microsoft can list and watch agents but documents no way to stop them. For agents Microsoft runs, listing, reading activity, naming, and refusing access at sign-in are documented. For agents on connected platforms, listing and reading activity are in preview and read-only, naming is announced but not in the documentation, refusing access at sign-in is marked sees only, in a tinted row, and confining or blocking the agent is not documented. For local agents on managed Windows devices, reading activity and confining or blocking are released, and listing, naming, and refusing access by device are in preview. The marks are Greyhound Research analysis.

Exhibit 1. On rivals’ platforms, Microsoft can see but not stop. Current through November 19, 2026. © Greyhound Research.

For agents Microsoft runs, Entra and Agent 365 can enforce. For local agents on managed Windows devices, Defender, Intune, and Windows containers can enforce, most of it in preview. For connected platforms, every control Microsoft documents is observation: a list, an activity view, and a name. The exhibit records what is written, not what the products may do later. It shows that the reach of Microsoft’s directory now runs well past the reach of its enforcement.

Greyhound Standpoint:

Greyhound Research’s position is that a name in a directory is an inventory, not a control, and that connected platforms should be valued as the first step in governing rivals’ agents rather than the last. The reason sits in Microsoft’s own documentation. The link to a rival platform is read-only, sync cannot add controls the platform lacks, and Agent 365 does not act on agents there. The power to stop a Salesforce agent stays with Salesforce and with whoever administers it.

That does not make the feature minor. One list of agents across rival platforms is something no CISO had before, and Microsoft deserves credit for building it openly against its competitors’ products. Security practice has long started with an asset inventory, because nothing can be protected until it has been found. The risk lies in how the list is read. Consolidated security consoles of earlier years were at times presented to boards as coverage when they offered visibility, and the difference tended to surface during an incident. This view rests on the documentation as it stood on November 19, and previews change. A Microsoft statement, platform by platform, of controls that reach the source runtime would change it.

At the next board risk review, security leaders should report connected agents as observed, not governed, and keep each platform’s own owners, controls, and means of stopping an agent in place.

The Deeper Read · About 3 minutes

Select a question to open the full answer

1What does the documentation say the connection does?

Read-only by design. Microsoft’s connected platforms documentation says connections are read-only and that registry sync cannot add capabilities the source platform does not already support. Agent 365 does not write to, change, or delete anything on the source platform, with one exception: an administrator may remove a synchronized agent if the customer has granted a separate, optional delete permission. Activity data is read from wherever the platform already writes its own logs.

What read-only tells a buyer. A read-only connection is a deliberate limit, and in a narrow sense a reassuring one: apart from the optional delete permission, Agent 365 cannot alter what runs on a rival’s platform. The same limit means it cannot correct anything there either. The documentation therefore answers two questions at once. It says the connection carries little risk to install, and it says the connection adds no new way to stop an agent. Both points belong in the pilot approval.

In the source’s own words

  • Microsoft Learn, connected platformsRegistry sync is read-only. It can't add capabilities that the source platform doesn't already support.Source
  • Microsoft Learn, connected platformsAgent 365 never invokes, executes, or interacts with agents in the source platform.Source
  • Microsoft Learn, connected platformsSynchronization, management, and observability capabilities vary by platform.Source
2Does a connected agent’s Entra Agent ID carry any enforcement?

Not stated. The Agent 365 post says Entra Agent ID lets organizations govern each agent’s lifecycle, apply supported Conditional Access and security policies, and revoke access. The connected platforms documentation does not describe an Agent ID being created for synchronized agents, or any policy reaching the source platform. Greyhound Research’s reading is that the word “supported” carries the weight. A Conditional Access policy acts when an agent asks Entra for access. A Salesforce agent working inside Salesforce does not ask Entra for anything.

Microsoft’s insider framing. Aleš Holeček, President and Chief Architect of Microsoft Security, wrote at Ignite that agents are the new insiders, and that discovered agents receive an Entra Agent ID and can be governed with Conditional Access. For agents Microsoft can see sign in, that holds. The companion Greyhound Research note on Azure agent infrastructure reached the same view from the other side: identity, not the machine, is where an agent’s authority is set. The companion note on confidential computing found the same of the sealed machine: the seal keeps the operator out, and the identity layer sets what the agent may do.

In the source’s own words

  • Microsoft Security Blog, Holeček at IgniteAgents are the new insiders.Source
  • Microsoft Security Blog, Holeček at IgniteDiscovered agents receive a Microsoft Entra Agent ID and can be governed with Conditional Access.Source
3How has Greyhound Research’s view of the agent control plane moved?

From the market to Microsoft’s own ground. In March 2026, Greyhound Research said there was no mature control plane governing how agents act, what they access, and how their actions can be reversed or contained. That was a view of the whole market. Since then, Agent 365, which carries Entra’s security features to agents, became generally available on May 1, and Ignite added sponsors, local agent IDs, and device controls. For agents Microsoft runs and for managed Windows devices, the view has therefore moved: a control plane now exists and is maturing, much of it still in preview. For agents on rivals’ platforms the March view still holds, because Microsoft’s documentation says the connection is read-only.

A test already on the record. In February 2026, Greyhound Research said that the most dangerous assumption in enterprise security is that valid identity implies safe behavior. This note applies the same test to the directory. A valid Entra entry shows who an agent is, not that anyone can stop it. A name settles authentication. The open question is authorization. Microsoft’s Ray Smith, Corporate Vice President for the Agent 365 product, described Agent 365 at Ignite as the observability platform for AI, one that turns shared visibility into management, governance, and security decisions. For connected agents, those decisions are still carried out on the platform where each agent runs.

In the source’s own words

  • Ray Smith, Microsoft, November 17Microsoft Agent 365 is the observability platform for AI, helping organizations turn shared visibilitySource

↑ Back to the section guide

Part 2The Plane And Its Limits

The Connected Roster: Eleven Platforms Named, Seven Documented

Greyhound SpotlightOpenAI, SAP, ServiceNow, and UiPath are named in Microsoft’s Ignite announcement and missing from the documentation’s list of supported platforms.

The roster matters because each name on it is a promise to a buyer who runs that platform. The announcement lists eleven. The documentation, which a buyer’s engineers will follow, lists seven, and for two of those, activity data is still marked coming soon. The other four may arrive by the December completion date. They are not documented today.

A roster is a commitment that ages. Each platform on it will change its own agent model, its logs, and its access rules on its own schedule, and any change can narrow what the connection shows. Integration catalogs in enterprise software have usually been judged at launch by their length and later by how well each entry was kept up. The same test applies here. Buyers who run one of the four undocumented platforms face a sharper version of it, since they cannot test what is not yet written. A point that gets less attention is cost: reading activity data may run queries the other provider bills for, so the roster can add charges to a rival’s invoice.

Table 3 sets each named platform against what the documentation says it supports. Synchronization means copying the platform’s list of agents into Agent 365. Observability means reading the agents’ activity.

Greyhound Research · Table 3 · The connected roster

Eleven platforms named, seven documented

Each platform Microsoft named for connected platforms at Ignite, against its documentation updated November 6, 2026.

Plan pilots on the seven documented platforms only, and treat the other four as announced until Microsoft documents them.

Named and documentedDocumented, activity coming soonNamed only
PlatformNamed at IgniteIn the documentationAgent list syncedActivity data
Salesforce AgentforceYesYesYesYes
Amazon Bedrock and AgentCoreYesYesYesYes
Google Vertex AIYesYesYesYes
Databricks GenieYesYesYesYes
Anthropic Claude Managed AgentsYesYesYesYes
Snowflake CortexYesYesYesComing soon
Oracle Generative AI AgentsYesYesYesComing soon
OpenAIYesNot listedNot documentedNot documented
SAPYesNot listedNot documentedNot documented
ServiceNowYesNot listedNot documentedNot documented
UiPathYesNot listedNot documentedNot documented
All seven documented: sync on request onlyStored key for every connectionDeleted agents may stay listed
Against this: Microsoft says the rollout completes in December, so the four named platforms may be documented by then.
Source: Microsoft Tech Community (Agent 365, November 17, 2026), Microsoft Learn (Connected platforms in Microsoft Agent 365, updated November 6, 2026).

Table 3 lists eleven platforms Microsoft named for connected platforms at Ignite, and shows seven in its documentation and four named only. Salesforce Agentforce, Amazon Bedrock and AgentCore, Google Vertex AI, Databricks Genie, and Anthropic Claude Managed Agents are documented with agent lists synced and activity data. Snowflake Cortex and Oracle Generative AI Agents are documented with agent lists synced and activity data coming soon. OpenAI, SAP, ServiceNow, and UiPath are named but not in the documentation. Notes beneath say all seven documented platforms sync only on request, use a stored key, and may keep agents deleted at the source.

Table 3. Eleven platforms named, seven documented. Data as of November 19, 2026. © Greyhound Research.

Seven platforms sync their agent lists, and five return activity data. The documentation adds limits that apply to all of them. Sync runs only when an administrator starts it. Agents deleted at the source may stay in Microsoft’s registry. Every connection uses a stored key. Each limit is ordinary for a preview, and each one is a gap between the registry and the truth that an identity team has to close by hand.

Greyhound Standpoint:

Greyhound Research holds that the connected roster is credible for inventory and not yet ready to run operations on. Seven documented platforms is a substantial start for a preview, and Microsoft has been candid about the limits. A registry that syncs only when an administrator asks, and that may keep agents deleted at the source, will drift from the platforms it mirrors between syncs. Each stored key is one more secret to vault, rotate, and revoke, and the duty to rotate it rests with the customer.

Identity teams have seen this before. Metadirectories and early provisioning connectors promised one view of accounts across systems, and the view was only as current as the last successful sync. The lesson of that period was that reconciliation is a standing process, not a project. The gap between eleven platforms named and seven documented is also typical of announcements that run ahead of documentation. It often closes, on a date the documentation sets rather than the announcement. The view would change with scheduled sync, removal of deleted agents, and keyless sign-in documented for each platform.

Identity leads should reconcile each platform’s agent list against Agent 365 at least monthly, and hold each connector key in a vault with a rotation date and read rights only. Buyers who run OpenAI, SAP, ServiceNow, or UiPath agents should obtain a documented date from Microsoft before the next pilot gate.

The Deeper Read · About 2 minutes

Select a question to open the full answer

1Which platforms does each Microsoft page name?

The announcement. The Agent 365 post says support is rolling out across Salesforce Agentforce, Amazon AgentCore, Google Vertex AI, Databricks, and Snowflake, and to Anthropic Claude Managed Agents, OpenAI, Oracle, SAP, ServiceNow, and UiPath, completing in December.

The documentation. Microsoft’s connected platforms page, updated November 6, lists Amazon Bedrock, Google Vertex AI, Salesforce Agentforce, Databricks Genie, Anthropic Claude Managed Agents, Oracle Generative AI Agents, and Snowflake Cortex. All seven sync. Oracle and Snowflake are marked coming soon for activity data, and activity collection is a Frontier preview, Microsoft’s early-access program.

In the source’s own words

  • Microsoft Tech Community, Agent 365 at IgniteSupport is rolling out across Salesforce Agentforce, Amazon AgentCore, Google Vertex AI, Databricks, Snowflake,Source
2What limits does the documentation state?

Four limits. The documentation says Agent 365 does not sync on a schedule, that removal of an agent deleted at the source is not guaranteed, that activity coverage varies by platform, and that sync adds no code-level capabilities. It also says reading activity data may run queries the other provider bills for, and that Frontier previews fall under the preview terms of the customer’s agreements.

Who holds the key. The same page assigns the creation, rotation, and revocation of each platform’s credentials to the platform’s administrator, with the Agent 365 administrator informed or coordinating. The key lives with Microsoft, encrypted. The duty to rotate it lives with the customer.

What the four limits tell a buyer. None of the four is unusual for a preview, and Microsoft has stated them plainly. Taken together, they describe the registry as a snapshot, refreshed on request, of platforms that each keep their own record. They also place two costs on the buyer that do not appear in Microsoft’s price: the other provider’s charges for the queries that read activity data, and the staff time to rotate each key. Both belong in the business case for a pilot.

In the source’s own words

  • Microsoft Learn, connected platformsAgent 365 doesn't currently sync on a schedule, and Microsoft doesn't prescribe a fixed synchronization interval.Source
  • Microsoft Learn, connected platformsDeletion reconciliation isn't guaranteed.Source
  • Microsoft Learn, connected platformsFrontier previews are subject to the existing preview terms of your customer agreementsSource

↑ Back to the section guide

Part 2The Plane And Its Limits

Where Enforcement Lives: One Agent, Two Rulebooks

Greyhound SpotlightMicrosoft documents a second, separate route for rivals’ agents: an agent on AWS or Google Cloud can sign in through Entra itself. Only that route puts Entra in a position to say no.

There are two ways a rival’s agent can enter Microsoft’s world, and they are easy to confuse. In the first, connected platforms copies the agent’s record into Agent 365. In the second, the agent’s own developer wires it to sign in through Entra, so Entra issues the pass the agent uses to reach a system. Only the second gives Entra a chance to refuse. An agent can be stopped wherever its pass is issued, and nowhere else.

The two routes can look alike on an Agent 365 screen, and that is the practical difficulty. A copied record and a federated identity both carry a name, yet only the second gives Entra a say when the agent acts. Federation has a long record in enterprise identity, and its value has rested on which systems agree to trust the pass it issues. Many companies, in Greyhound Research’s expectation, will end up with a mix, a few agents federated and most simply synced, because federation depends on each agent’s developer doing the wiring. That makes the route an architecture decision to be taken agent by agent, and recorded, rather than a setting switched on once for the whole estate.

Exhibit 2 follows one agent on a rival’s platform through five steps, from connection to action, and marks which steps Microsoft has written and where the pass is issued.

Greyhound Research · Exhibit 2 · One connected agent, step by step

Microsoft has written the listing, not the stopping

Five steps for one agent on a rival platform, from connection to action, read against Microsoft’s documentation and announcement, November 19, 2026.

The step that decides harm is the last one, and it sits with whoever issues the agent’s credentials.

Written by MicrosoftPartly writtenNot written
1Connect WrittenAn administrator links the rival platform to Agent 365Written. The documentation says each connection uses a stored key held encrypted by Microsoft.Open. The announcement says keyless federation instead.
2List WrittenAgent 365 copies the platform’s agents into its registryWritten. Read-only, with no change made on the rival platform.Open. Sync runs only on request, and deleted agents may stay listed.
3Name Partly writtenEach agent receives an Entra Agent IDWritten. The announcement says connected agents receive one automatically.Open. The documentation does not describe the ID being created.
4Watch Partly writtenAgent 365 reads what the agent doesWritten. Activity read from the platform’s own logs, for five of seven platforms.Open. A Frontier preview, with coverage that varies.
5Act Not writtenThe agent does its work under its own platform’s credentialsWritten. None. Microsoft documents that Agent 365 never acts on the source platform.Open. Who stops the agent, and how fast.
Against this: Microsoft documents a separate route in which an agent on AWS or Google Cloud signs in through Entra itself, which lets Entra refuse it.
Source: Greyhound Research analysis of Microsoft Learn (connected platforms, integration paths for agents) and Microsoft’s Agent 365 post (November 17, 2026). The steps and grades are Greyhound Research’s.

Exhibit 2 follows one agent on a rival platform through five steps, and shows Microsoft has written the listing and not the stopping. Step one, connect, is written: each connection uses a stored key, while the announcement says keyless federation. Step two, list, is written: read-only, synced on request, and deleted agents may stay listed. Step three, name, is partly written: the announcement says each agent receives an Entra Agent ID and the documentation does not describe it. Step four, watch, is partly written: activity is read for five of seven platforms in a Frontier preview. Step five, act, is not written: the agent works under its own platform’s credentials, and Microsoft documents that Agent 365 never acts on the source platform. The steps and grades are Greyhound Research analysis.

Exhibit 2. Microsoft has written the listing, not the stopping. Current through November 19, 2026. © Greyhound Research.

Microsoft has written the first steps well: connecting, listing, and naming. The step that decides harm is the last one, when the agent acts. For an agent that only syncs, that step happens under the rival platform’s rules and credentials. For an agent wired to Entra, Entra can refuse the pass, but only for systems that trust Entra’s passes and only for the agent’s own access, not for actions taken as a signed-in user. Greyhound Research’s reading is that the enforcement map follows the pass, not the directory.

Greyhound Standpoint:

As Greyhound Research sees it, the most useful question a CISO can ask of any agent is not whether it is in the directory but who issues its credentials, because that party can refuse the agent at the moment it acts. A copied record gives Microsoft no say. Federation through Entra puts Entra in the path of the agent’s own access, but only for systems that accept Entra’s passes, and only for the agent acting as itself rather than for a signed-in user.

Federated identity has long worked this way. It extended reach without moving the authority of each system that relies on the pass, and agents inherit that design. Microsoft’s sponsor model, which passes sponsorship to a manager when a sponsor leaves, is a practical answer to the ownerless agent. The limit of this reading is that the guide does not say how far Entra’s controls reach beyond Microsoft’s systems, so the view rests on how passes work rather than on a Microsoft statement. A documented list of systems where an Entra refusal stops a rival’s agent would sharpen it.

Identity leads should draw a two-column map for every agent this quarter, where it is listed and who issues its credentials, and the incident plan should name, for each mismatch, the platform and person who can revoke it. This agrees with the companion note’s advice to treat every agent as a privileged account.

The Deeper Read · About 2 minutes

Select a question to open the full answer

1How does an agent on a rival’s platform sign in through Entra?

Integration paths. Microsoft’s integration guide, updated November 17, lists routes for agents outside Microsoft. An agent in a container on Amazon Bedrock can run Microsoft’s authentication software alongside it. An agent on AWS or Google Cloud can use that cloud’s own workload identity, through federation, for access as itself. The guide says federation supports only the agent’s own access, not calls on behalf of a signed-in user.

Who the pass is good for. The guide does not say whether Entra’s controls reach systems outside Microsoft. Greyhound Research’s reading is that Entra can refuse a pass only for systems that accept Entra’s passes, which includes Microsoft 365, Azure, and any application the customer has set up to trust Entra.

What the guide’s limit tells a buyer. The guide is careful to say that federation covers only the agent acting as itself. Agents often act on behalf of a signed-in user, with that user’s permissions, and for those calls the guide offers no route through federation. Read with the rest of the guide, that means an Entra refusal reaches part of an agent’s activity, not all of it. Buyers mapping where each agent can be stopped should record the two kinds of access separately.

In the source’s own words

  • Microsoft Learn, integration paths for agentsThe agent runs in a container on Amazon Bedrock and uses the Auth SDK sidecar.Source
  • Microsoft Learn, integration paths for agentsWorkload identity federation exchanges a token from an external identity provider for a Microsoft Entra token.Source
  • Microsoft Learn, integration paths for agentsSupports app-only access. It doesn't authenticate users or support calls as a signed-in user or an agent's user account.Source
  • Microsoft Learn, What is Microsoft Entra Agent IDMicrosoft Entra Agent ID works with agents built on Microsoft and non-Microsoft platforms.Source
2What governance does Entra add once an agent signs in through it?

Sponsors and expiry. Entra ID Governance lets each agent identity have a sponsor, a person accountable for its lifecycle and access, and moves sponsorship to the sponsor’s manager when the sponsor leaves. Access granted through an access package can expire on a set date if the sponsor does not renew it. Governance for agents requires Microsoft 365 E7, or Agent 365 with at least Entra P1 or Microsoft 365 E3.

Task-scoped access. Holeček wrote that Entra ID Governance dynamic permissions grant task-scoped, time-bound access, and that these remain in limited preview. That is the finer rule the companion note on Azure agent infrastructure asked for, and it is not yet released.

In the source’s own words

  • Microsoft Learn, governing agent identitiesSponsors of agent identities are human users accountable for making decisions about its lifecycle and access.Source
  • Microsoft Learn, governing agent identitiesIf the sponsor is leaving the organization, sponsorship of the agent identities can be automatically transferred to their manager.Source
  • Microsoft Security Blog, Holeček at IgniteMicrosoft Entra ID Governance dynamic permissions grant task-scoped, time-bound accessSource

↑ Back to the section guide

Part 3Local Agents And Liability

Local And Shadow Agents: The Laptop Becomes The Checkpoint

Greyhound SpotlightMicrosoft’s own documentation warns that, during preview, Claude Code might keep processing a tool’s output even after Defender reports a block.

The agents that worry security teams most are often not in any cloud. They run on a developer’s laptop, with the developer’s access, and nobody approved them. Here Microsoft can do more than watch, because it already manages the device. Defender can find the agent, Entra can give it an ID, Windows can confine it to a container, and Intune can cut the device off from company data if an unapproved agent is running.

This is the one area where agents arrive on ground Microsoft already governs. Endpoint management has matured over many years of patching, device compliance, and application control, and most large Windows estates already run the tools involved. That gives endpoint teams a head start that cloud identity teams do not have. What is new is the kind of software being governed. A coding agent works with its developer’s access and can change files, call tools, and send data quickly. Greyhound Research expects developer groups to feel these controls first, and the tension between developer speed and device policy, familiar from earlier waves of unapproved software, to return in a sharper form.

Exhibit 3 follows one local agent through six controls, from discovery to network control, and marks the status of each and the gap Microsoft itself records.

Greyhound Research · Exhibit 3 · Where a local agent meets Microsoft’s controls

On the device, Microsoft can stop agents, with one gap

One local agent, such as Claude Code or GitHub Copilot CLI, on a managed Windows device, along the six controls Microsoft announced or released, November 19, 2026.

The device is where Microsoft can stop agents today. Contain first, and keep inspection in audit mode until it blocks before a tool runs.

ReleasedPreviewReleased with a documented gap
Find and nameReleased, then previewDefender finds more than 60 agent types. Agent IDs start with Copilot CLI and Claude Code.
ContainReleasedExecution Containers enforce file and network limits the agent cannot widen.
InspectReleased on Windows, with a gapA check after a tool responds cannot undo what the tool did.
Block and routePreviewIntune and Conditional Access cut the device off. Global Secure Access controls its traffic.
Against this: Microsoft advises starting runtime protection in audit mode, and Execution Containers limit what an agent can reach whatever inspection finds.
Source: Greyhound Research analysis of Microsoft’s Agent 365 post, the Book of News, the Windows Experience Blog (November 17, 2026), and Microsoft Learn (AI agent runtime protection, updated November 11, 2026).

Exhibit 3 follows one local agent along a bar of six controls from start to data: discover, name, contain, inspect, block the device, and control the network. Discovery and containment are released. Naming is in preview. Inspection is released on Windows per Microsoft’s announcement, marked with a gap, because Microsoft’s documentation records that a check after a tool responds cannot undo what the tool did. Blocking the device through Intune and Conditional Access and controlling network traffic through Global Secure Access are in preview. A strip beneath groups the steps into four. The mapping is Greyhound Research analysis.

Exhibit 3. On the device, Microsoft can stop agents, with one gap. Current through November 19, 2026. © Greyhound Research.

Two controls are generally available: finding the agent and inspecting it while it runs on Windows. Confinement in Microsoft Execution Containers has been released since October. Naming the agent, blocking the device, and controlling its network traffic are in preview. The gap Microsoft records sits in the middle of the path: a check made after a tool runs cannot undo what the tool already did. Of every area in this note, the device is where Microsoft’s enforcement is strongest.

Greyhound Standpoint:

Greyhound Research’s judgment is that the managed device is where Microsoft’s agent controls are strongest, because Microsoft already owns the tools that can stop a local agent. Defender finds and inspects it, Windows contains it, and Intune can cut the device off from company data. Discovery of more than sixty agent types and containment the agent cannot widen are solid foundations.

One documented gap shapes how far to rely on them. Runtime inspection checks content as it enters or leaves the agent’s reasoning, so a check made after a tool responds cannot undo what the tool did. Endpoint security has followed this path before. Detection products generally began in audit mode and moved to blocking as trust grew, while prevention came from limiting what software could reach. Microsoft deserves credit for publishing the gap rather than leaving customers to find it. The view would change if runtime protection documented a block before a tool runs for every supported agent, and if its status read the same on every Microsoft page.

Endpoint leads should set container boundaries through Intune before agents receive IDs, block unapproved agents first, and keep runtime protection in audit mode until the pilot gate shows its accuracy. Agents that can delete or send data should run with narrow rights whatever the inspection reports, and the switch to block mode should be signed off by the CISO.

The Deeper Read · About 3 minutes

Select a question to open the full answer

1What does each local control do?

Discovery and naming. The Agent 365 post says Defender and Global Secure Access now discover more than 60 local AI agents, that Defender’s discovery is generally available, and that Defender supports managed Windows and macOS devices. Entra Agent ID assigns an identity to supported local agents Defender discovers, starting with GitHub Copilot CLI and Claude Code, in preview. Greyhound Research found in May 2025 that 67 percent of IT leaders surveyed for Greyhound CIO Pulse 2025 saw agent execution at the edge as the greatest blind spot in their AI risk posture.

Containment. Pavan Davuluri, Executive Vice President of Windows + Devices at Microsoft, wrote at Ignite that Microsoft Execution Containers have been generally available on Windows 11 since October. Intune policies set the file and network boundaries they enforce. The Agent 365 post says those boundaries are enforced outside the agent’s control. Local Micro VMs, a stronger hardware-backed boundary, are experimental.

In the source’s own words

  • Microsoft Tech Community, Agent 365 at IgniteDefender supports managed Windows and macOS devices and ties each agent to its device and userSource
  • Microsoft Tech Community, Agent 365 at IgniteThose boundaries are enforced outside the agent's control, so the agent cannot grant itself more access.Source
  • Windows Experience Blog, Davuluri at IgniteMicrosoft Execution Containers (MXC) has been generally available on Windows 11 since October.Source
  • Greyhound Research, Build 2025 open agentic web notenow believe agent execution at the edgeSource
2What are runtime protection’s documented limits?

Audit first. Defender’s runtime protection page says Defender inspects points where content enters or leaves the agent’s reasoning, and recommends starting in audit mode before switching to block mode. It names Claude Code, Codex CLI, and GitHub Copilot CLI as agents that expose the needed hooks. Network inspection does not support agents that use certificate pinning or HTTP/3. The Agent 365 post says Defender can block malicious instructions before the agent acts. That addresses a risk an identity cannot: poisoned input that turns a correctly named agent against its owner.

After the fact. The page records a known issue: during preview, Claude Code might keep processing a tool response in block mode after Defender reports a block, and a check made after a tool responds cannot undo what the tool already completed. The Book of News says runtime protection will be available to Defender for Endpoint P2 customers.

What the known issue tells a buyer. The issue marks the line between inspection and prevention. Inspection can stop the next step. It cannot reverse the last one. For agents that read code or summarize files, that may be acceptable during a pilot. For agents that can delete records or send data outside the company, the protection that counts is the narrow right granted before the agent starts, which is why containment and permissions carry more weight than inspection in this note’s reading.

In the source’s own words

  • Microsoft Learn, AI agent runtime protectionDuring preview, Claude Code might continue processing a tool response even when AI agent runtime protection is configured in Block mode and Defender reports a post-tool-response block.Source
  • Microsoft Tech Community, Agent 365 at Ignitecan block malicious instructions before the agent acts.Source
  • Microsoft Learn, AI agent runtime protectionstarting in audit mode to observe detections and validate accuracy before switching to block mode.Source
  • Microsoft’s Book of News, DefenderRuntime protection for local AI agents will be available to Defender for Endpoint P2 customers.Source
3What does Intune add?

Block the device. The Book of News says Intune will let IT discover which agents run on managed devices, mark devices with unapproved agents as noncompliant, and apply Conditional Access to block those devices from company data. For approved agents, Intune policies will define what they can reach and do inside Execution Containers. All of it is in preview.

In the source’s own words

  • Microsoft’s Book of News, Intunemark devices with unapproved agents as noncompliant with Intune software compliance policiesSource
  • Microsoft’s Book of News, Intuneapply Microsoft Entra Conditional Access to block those devices from company data.Source

↑ Back to the section guide

Part 3Local Agents And Liability

When Enforcement Fails: Who Answers For A Rival’s Agent

Greyhound SpotlightMicrosoft’s responsibility table for connected platforms gives the platform’s administrator ownership of creating, granting, rotating, and revoking credentials. Agent 365’s administrator is informed.

A single directory invites a single assumption: if the agent is in Microsoft’s list, Microsoft is watching it. When a connected agent leaks data or takes a wrong action, the questions are who should have stopped it and who pays. Microsoft’s documents answer the first question by design. Agent 365 does not act on rival platforms, so it is not placed to be the party that stops the agent there.

Questions of liability tend to be settled after the first serious incident, and buyers are better placed when they settle them before it. Microsoft’s documents already set out its side in plain terms, so the open work sits with the customer and each platform owner. What changes with agents is the clock. An agent can take many actions in the time two teams need to agree who owns the response, so a revocation duty with no time attached may offer little protection. The less visible consequence concerns evidence. When something goes wrong, the logs that count sit on the rival’s platform, while the first sign may appear on a Microsoft screen, and the incident plan has to join the two.

Exhibit 4 sorts the duties into three groups: what Microsoft has written, what each platform owner holds, and what the customer needs to write down because no document yet assigns it.

Greyhound Research · Exhibit 4 · Who answers when enforcement fails

Visibility moves to Microsoft, liability stays put

Duties around an agent on a connected platform, as Microsoft’s documentation assigns them and as they remain to be written, November 19, 2026.

Greyhound Research’s judgment: a listing moves visibility, not liability. Write the duty to stop an agent into each platform’s contract.

Written by MicrosoftA narrow, read-only roleRead-only. Agent 365 never invokes or interacts with agents on the source platform.No new controls. Sync cannot add what the platform does not support.Preview terms. Frontier previews fall under the customer’s preview terms.Written
Held by the platform ownerThe means to stop the agentCredentials. Creates, grants, rotates, and revokes them.Permissions. Sets what the agent may do on the platform.Logs. Keeps the activity record Agent 365 reads.Owned by the platform
For the customer to writeDuties no one has assignedRevocation time. How fast an agent must be stopped.Notice. Who is told, by whom, and when.Evidence. Which log is the record in a dispute.Not yet assigned
Against this: Microsoft’s narrow role is stated plainly in its documentation, which gives buyers what they need to assign the rest.
Source: Greyhound Research analysis of Microsoft Learn (Connected platforms in Microsoft Agent 365, updated November 6, 2026). The grouping is Greyhound Research’s. Not legal advice.

Exhibit 4 groups the duties around an agent on a connected platform into three cards. Written by Microsoft: a read-only role, in which Agent 365 never invokes or interacts with agents on the source platform, sync cannot add controls the platform lacks, and Frontier previews fall under preview terms. Held by the platform owner: the credentials, the agent’s permissions, and the activity logs. For the customer to write: how fast an agent must be stopped, who is told, and which log is the record. The grouping is Greyhound Research analysis and not legal advice.

Exhibit 4. Visibility moves to Microsoft, liability stays put. Current through November 19, 2026. © Greyhound Research.

Microsoft has written its role narrowly and clearly: read-only, no action on the source platform, preview terms. The platform owner holds the credentials, the agent’s permissions, and the means to stop it. The customer holds everything in between, including how fast a revocation must happen, who is told, and which log counts as evidence. The gap is real, but it is not Microsoft’s to close alone. This is a reading of published documents, not legal advice.

Greyhound Standpoint:

Greyhound Research’s reading is that connected platforms moves visibility to Microsoft and leaves liability where it was, with each platform and with the customer who set the agent’s rights. Microsoft has written its role narrowly and clearly: read-only, no action on the source platform, and preview terms. That clarity is a strength. It also means a buyer who treats Agent 365’s list as cover is relying on an assurance no contract gives.

The structure resembles multi-vendor outsourcing, where one service management layer sat above several providers. Reporting was consolidated, yet each provider’s contract still set what it had to fix and how fast, and disputes tended to gather where no contract named an owner. Agent programs are taking the same shape, with Microsoft as the reporting layer and each platform as the operator. The limit of this view is that it rests on Microsoft’s documentation, not on any contract or terms of service, which were not read for this note. A written statement of Microsoft’s role in Agent 365’s terms would refine it, and this note asks Microsoft for one.

At each platform’s next renewal, general counsel and procurement should write the enforcement duty into the contract: who can stop an agent, within what time, and whose log is the record. The Agent 365 order should state that connected agents are observed only during preview, and each term should be confirmed with counsel.

The Deeper Read · About 1 minute

Select a question to open the full answer

1What does Microsoft’s documentation commit to?

A narrow role. The connected platforms documentation says Agent 365 never invokes or interacts with agents on the source platform, does not write to it, and reads activity from logs the platform already keeps. Its responsibility table gives the platform administrator ownership of credentials and permissions. Frontier previews fall under the preview terms of the customer’s agreements, and the documentation makes no statement about the third party’s own terms.

Credentials Microsoft holds. The same page says Agent 365 keeps the credentials entered for a connection encrypted inside Microsoft 365. That makes the connector key a Microsoft-held secret with a customer-owned rotation duty, which the incident plan should name.

What the responsibility table tells a buyer. The table is clear and useful, and it is also a list of duties the customer inherits. Each row that gives the platform administrator ownership implies a person on the customer’s side who must hold that role, for every connected platform. In many companies that person sits outside the security team, in the group that runs the business application. Naming those people, and the time each has to act, turns the table from a description into an operating plan.

In the source’s own words

  • Microsoft Learn, connected platformsAgent 365 never invokes, executes, or interacts with agents in the source platform.Source
2What has Greyhound Research said about terms that fall short?

Consolidation with conditions. Greyhound Research’s November 15 note on the Microsoft 365 outage held that consolidating on one suite remains right while its standard terms fall short. The same logic applies here. A single agent directory is the right design, and its terms should say what it does not do. The July 2026 note on agent runtimes set four tests for any agentic application, and the fourth, reversing a bad chain of actions with a named owner, sits on the platform where the agent acts.

In the source’s own words

  • Greyhound Research, July 2026 agent runtime noteA buyer who cannot do these four things does not have an agentic applicationSource

↑ Back to the section guide

Part 4The Platform Contest

Every Vendor Wants The Directory: A Year Of Rival Bids

Greyhound SpotlightAWS’s documentation says its agent registry works with agents deployed on AWS, on premises, or on any other cloud. Microsoft is not alone in bidding for the cross-vendor list.

Ignite’s announcement lands in a contest. Through 2026, every company that sells agents or identity has offered to be the place where all agents are listed. AWS, Google, Okta, and ServiceNow have each made the offer in some form. A company is unlikely to run five master lists well, so the contest is over which list becomes the record and which ones feed it.

The contest will be decided less by announcements than by what is already installed, what connects most cheaply, and which team owns the budget. On those terms Microsoft starts from a strong position in most large enterprises, while AWS and Google start from where many of their customers’ agents run. Greyhound Research expects most buyers to live with overlapping registries for some time, and the cost of that overlap will show up as reconciliation work rather than as a line on an invoice. Buyers also gain leverage while the contest is open. Every vendor in it has reason to make its registry easier to feed, easier to export from, and easier to connect to the others.

Exhibit 5 places the main moves of 2026 on a month grid, from Microsoft’s March pricing to Ignite’s connected platforms.

Greyhound Research · Exhibit 5 · The 2026 contest calendar

Every vendor bid for the agent list in 2026

Moves by Microsoft, AWS, ServiceNow, and Okta to list or govern agents beyond their own platforms, by month, 2026.

The contest is unlikely to settle soon. Name one directory of record now, and let the other registries feed it.

A bid for the cross-vendor agent list fell in this monthDecember is a scheduled date, not an event
JanQuiet
FebQuiet
MarMicrosoft, AWSAgent 365 dated and priced. AWS tool-call rules released
AprAWSAgent Registry enters preview, April 9
MayMicrosoft, ServiceNowAgent 365 released May 1. AI Control Tower reaches beyond ServiceNow
JunQuiet
JulQuiet
AugAWS, ServiceNowAWS Agent Registry released August 31. AI Control Tower release expected
SepOktaImports agents from Bedrock and Agentforce
OctMicrosoftNadella cites 50 million agents, October 28
NovMicrosoftConnected platforms at Ignite, November 17
DecMicrosoftConnected platform rollout due to complete
AWS: Agent Registry generally available, documentation names any cloudGoogle: cross-cloud registration in limited preview
Against this: Each rival bids from its own ground, and only Microsoft’s list sits beside the workforce directory most companies run.
Source: Official Microsoft Blog (March 9, 2026), AWS What’s New (March 3, April 9, and August 31, 2026), ServiceNow (May 5, 2026), Okta Newsroom (September 22, 2026), Satya Nadella (October 28, 2026), Microsoft Tech Community (November 17, 2026), AWS and Google Cloud Documentation.

Exhibit 5 places bids for the cross-vendor agent list on a 2026 month grid, and shows eight months with a move. In March, Microsoft dated and priced Agent 365 and AWS released rules on each agent tool call. In April, AWS put Agent Registry in preview. In May, Agent 365 was released on May 1 and ServiceNow said AI Control Tower reaches beyond ServiceNow. In August, AWS released Agent Registry and ServiceNow expected AI Control Tower’s release. In September, Okta said it imports agents from Bedrock and Agentforce. In October, Nadella cited 50 million agents in Agent 365. In November, Microsoft announced connected platforms at Ignite. December is the scheduled completion of the rollout. Notes beneath say AWS’s Agent Registry is generally available, with documentation that names any cloud, and Google’s cross-cloud registration is in limited preview.

Exhibit 5. Every vendor bid for the agent list in 2026. Data as of November 19, 2026. © Greyhound Research.

The bids cluster in the spring and the autumn. Each vendor starts from its own ground: AWS and Google from their agent runtimes, Okta from sign-in, ServiceNow from IT workflow, and Microsoft from Entra, the directory that already holds most companies’ employees. That last point is Microsoft’s advantage. Its list of agents sits next to the list of the people who sponsor them. On release status, AWS is ahead: its Agent Registry has been generally available since August 31, while Microsoft’s connected platforms are in preview.

Greyhound Standpoint:

Greyhound Research concludes that Microsoft has the strongest claim to be the directory of record for agents, because Entra already holds the people who sponsor and answer for each one, while rivals keep the stronger claim to enforce on their own runtimes. Buyers are better served by assigning each role than by choosing one winner.

The deciding link is the sponsor. An agent needs a person who answers for it, and that link stays current only if the directory also knows when the person changes role or leaves. Rivals bring real strengths in runtime control, sign-in, and IT workflow, but none starts from the employee record as Entra does for most Microsoft customers. Consolidation cycles in enterprise software have tended to settle on one system of record that others feed, rather than one product doing everything. Microsoft also lists its own agents beside its rivals’, so it warrants the scrutiny due to any vendor that governs what it sells, and the record should survive a change of directory. The view would weaken if a rival tied its registry to the workforce record as closely, and AWS’s re:Invent is the next test.

Enterprise architects should name one directory of record for agents at the next architecture review, let the other registries feed it, and keep enforcement on each runtime. They should favor the directory that can show, platform by platform, what it enforces as well as what it lists.

The Deeper Read · About 2 minutes

Select a question to open the full answer

1What has each rival offered?

AWS. AWS documents AgentCore Identity as an agent identity service compatible with existing identity providers, and its 2025 launch post named Microsoft Entra ID among them. Its documentation says AgentCore Registry works with agents and tools deployed on AWS, on premises, or on any other cloud. AWS Agent Registry entered preview on April 9, 2026, and became generally available on August 31, 2026, as a governed catalog of agents, tools, skills, MCP servers, and custom resources within an organization.

Google. Google’s Agent Registry page describes a central hub for governance, discovery, and search. Registration of agents on Kubernetes clusters outside Google Cloud is in limited preview.

Okta and ServiceNow. Okta said on September 22 that known agents can be imported from platforms like AWS Bedrock or Salesforce Agentforce, and that it will surface unmanaged agents on employee devices. ServiceNow said on May 5 that AI Control Tower discovers AI assets including systems beyond ServiceNow, with general availability then expected in August 2026.

In the source’s own words

  • AWS Documentation, AgentCore overviewAgentCore Registry works with any MCP Server, Agent, Skill or Custom Resource, deployed on AWS, On-Prem or on any other Cloud environment.Source
  • AWS Machine Learning Blog, August 15, 2025For example, Amazon Cognito, Okta, and Microsoft Entra ID, among others.Source
  • Okta Newsroom, September 22surfaces unmanaged agents running on employee devices before they become blind spots.Source
  • ServiceNow, May 5AI Control Tower enhancements enter Innovation Lab in May with general availability expected in August 2026.Source
2Why does the directory’s link to people matter?

The sponsor is a person. Entra’s governance model ties each agent to a human sponsor and hands sponsorship to a manager when that person leaves. Only a directory that also holds the employees can do that without a second feed. Greyhound Research wrote in May 2025 that Entra Agent ID “isn’t just a feature,” and called it a policy control plane for agent-based computing. That holds where Entra issues an agent’s credentials. Ignite extends Entra’s list, not its policy, to rivals’ platforms, so for connected agents the 2025 description fits the directory and not yet the control.

What the sponsor link tells a buyer. The sponsor is the most practical feature in this contest, because it answers the question every audit asks first: who owns this? A registry that cannot keep its owners current becomes a list of orphans within a few reorganizations. Entra’s advantage comes from the employee records that many companies already feed into it, and a rival would need a comparable feed to match it. Buyers comparing registries should test how each handles a sponsor who leaves before they test anything else.

In the source’s own words

  • Greyhound Research, Build 2025 agents noteEntra Agent ID isn't just a featureSource

↑ Back to the section guide

Part 4The Platform Contest

Three Clouds Compared: Microsoft Leads On Reach, AWS On Enforcement

Greyhound SpotlightAWS has checked each agent’s tool requests against written rules, as a generally available feature, since March 3, 2026. Microsoft’s equivalent at Ignite, the MCP Firewall, is in preview.

Each of the three large clouds now gives every agent an identity of its own. They differ in how far that identity reaches and how firmly it is enforced. Microsoft reaches furthest, across rivals’ platforms and employees’ computers. AWS enforces most firmly at the moment an agent calls a tool. Google builds its agent identity on SPIFFE, an open standard for workload identity.

Few large companies run agents on one cloud alone, which is why the comparison matters. A company with Microsoft 365 for its staff, AWS for its applications, and Google for its data work will meet all three identity models in the same quarter. Greyhound Research expects the differences to narrow over the coming year as each vendor fills the gaps the others have exposed, and it would be unusual for any lead in this table to last unchanged. The practical question is which layer the company keeps for itself. The point most comparisons miss is that an open standard such as SPIFFE may matter as much as any feature, because it can decide how easily agent identities move between vendors.

Table 4 compares the three layer by layer on each vendor’s own documents.

Greyhound Research · Table 4 · Three clouds, layer by layer

Microsoft leads on reach, AWS on enforcement

Microsoft, AWS, and Google compared on agent identity and governance, on each vendor’s own documents, November 19, 2026.

No vendor leads on both reach and enforcement. List agents in the directory, and keep tool-level rules on each runtime.

Generally available or documentedPreviewNot statedNot assessed or unlabeled
LayerMicrosoftAWSGoogle
Agent identityAvailable Entra Agent ID, for all Entra customersGenerally available AgentCore Identity, with AgentCore since October 2025No label on page Agent Identity, built on SPIFFE
Works with other identity providersIs the provider Entra itselfDocumented Names Entra ID, Okta, and CognitoNot assessed
Registry beyond its own platformPreview Eleven platforms named, read-onlyGenerally available Agent Registry, since August 31, 2026Documentation names any cloudLimited preview Kubernetes outside Google Cloud
Rules on each tool callPreview MCP FirewallGenerally available Policy, since March 3, 2026No single label Agent Gateway, tool-level rules, deny by default
Agents on employee devicesGenerally available Discovery and containmentNaming in previewNot assessedNot assessed
Enforcement on a rival’s platformNot statedNot statedNot stated
Against this: Microsoft holds the workforce directory, the endpoint tools, and Microsoft 365, which no rival read for this note combines.
Source: Microsoft Book of News, Learn, and Agent 365 post (November 17, 2026), AWS What’s New (October 13, 2025, and March 3 and August 31, 2026), AWS Documentation and Machine Learning Blog (August 15, 2025), Google Cloud Documentation (updated November 17 and 18, 2026).

Table 4 compares Microsoft, AWS, and Google on six layers of agent identity and governance, and shows Microsoft ahead on reach, AWS ahead on rules at each tool call, and no vendor stating enforcement on a rival’s platform. Agent identity is available on Microsoft, generally available on AWS, and unlabeled on Google’s SPIFFE-based page. AWS documents compatibility with Entra ID, Okta, and Cognito. Registries beyond each vendor’s platform are in preview on Microsoft, generally available on AWS since August 31, 2026, with documentation that names any cloud, and in limited preview on Google. Rules on each tool call are generally available on AWS since March 3, 2026, in preview on Microsoft, and documented on Google’s Agent Gateway without a single release label. Agents on employee devices are assessed only for Microsoft. Several cells are not assessed.

Table 4. Microsoft leads on reach, AWS on enforcement. Data as of November 19, 2026. © Greyhound Research.

Microsoft is ahead on reach, with connected platforms, local agents, and the workforce directory. AWS is ahead on enforcement at the tool call, as the companion note on Azure agent infrastructure also found. Google’s identity rests on an open standard, and its gateway denies by default and documents rules at the tool-request level, with no single release label. No vendor documents enforcement over agents running on a rival’s platform. Outside the endpoint row, one cell is not assessed. The table does not rank the three on endpoints, where only Microsoft was read.

Greyhound Standpoint:

Greyhound Research’s assessment is that Microsoft leads on reach and AWS leads on enforcement at the tool call, and that neither lead settles the contest, because the two are different jobs. A directory that lists every agent answers what exists and who sponsors it. A gateway that checks every tool request answers what an agent may do at the moment it acts. Most companies will need both.

AWS has offered generally available rules on each tool request since March 3, 2026. Google documents tool-level rules in its Agent Gateway, which denies by default, though the page carries no single release label. Microsoft’s MCP Firewall is in preview. Microsoft’s advantage is its identity estate, across rivals’ platforms, devices, and the workforce directory. Platform shifts of this kind have tended to split the market before settling it. In earlier cloud waves, many buyers paired one provider’s directory with another’s runtime controls, with a policy layer they owned in between, and that kept options open while vendors closed their gaps. The view would change if Microsoft released its tool-level rules with documentation, or if any vendor documented enforcement on a rival’s platform.

Enterprise architects should keep tool-level rules on each runtime or in a gateway the company controls, as the companion note advised, and use the directory to list agents and their sponsors. The decision belongs at the next architecture review, before any runtime’s own rules are retired.

The Deeper Read · About 1 minute

Select a question to open the full answer

1What does each vendor document?

AWS. AWS made AgentCore generally available on October 13, 2025. Policy in AgentCore became generally available on March 3, 2026, with a gateway that intercepts agent-tool traffic and evaluates each request against Cedar rules before allowing it.

Google. Google’s Agent Identity page gives each agent a cryptographic identity based on SPIFFE, with no release label on the page. Its Agent Gateway page, updated November 18, says all connections are blocked unless a policy grants access, and that it reads tool names from requests to servers built on the Model Context Protocol. Cross-cloud use is in limited preview.

In the source’s own words

  • AWS What’s New, March 3, 2026intercepts agent-tool traffic and evaluates each request against the policiesSource
  • Google Cloud Documentation, Agent Identity overviewAgent Identity provides a strongly attested, cryptographic identity for each agent that is based on the SPIFFE standard.Source
  • Google Cloud Documentation, Agent Gateway overviewBy default, all connections are blocked unless an explicit IAM policy grants access.Source
2Where does Microsoft’s tool-level control stand?

In preview. The Book of News says the MCP Firewall will control access to MCP servers, tools, resources, prompts, protocols, and transports, in preview. The companion note on Azure agent infrastructure found that Agent Grid, the policy layer for the new Azure sandboxes, had no documentation through 02:00 UTC on November 18. Microsoft’s identity estate is its advantage. Its tool-level rules are still to be released.

What the release gap tells a buyer. A preview of tool-level rules is not a weakness in itself. It does mean that, for now, an agent estate centered on Microsoft relies on rules held elsewhere, in AWS, in Google, or in a gateway the company runs. Buyers can expect the gap to narrow, and should avoid building plans around the date it closes. The safer design keeps tool-level rules in a layer the company can move between vendors, so that whichever release lands first can be adopted without rework.

↑ Back to the section guide

Part 5The Risks And The Counter-Argument

Risks To Watch: Where A Cross-Vendor Plane Can Fail

Greyhound SpotlightEach of the five risks shows first as a mismatch between two lists the buyer already holds.

Greyhound Research sees five risks in using one directory for agents across vendors. Most arise because a list feels like control. Each shows first inside the buyer’s own records, where it can be caught early.

These risks behave differently from the ones a vendor review usually looks for. None needs Microsoft to suffer an outage or a breach, and none would be covered by a service credit, because each sits in the buyer’s own records. Greyhound Research expects the stale registry and the standing keys to show first, since both appear as soon as two lists are compared, and the liability gap last, since it tends to surface only when an incident tests it. That sequence is useful in practice. Risk owners can run the cheap checks this quarter and carry what they find into the contract work at each platform’s renewal, where evidence carries more weight than a general concern.

Table 5 sets each risk beside its early warning and its owner. Every early warning can be checked without help from Microsoft.

Greyhound Research · Table 5 · Risk register

Every risk shows first as two lists that disagree

Five risks in using one directory for agents across vendors, each with its early warning and its owner.

Every warning sits in the buyer’s own records. Check all five this quarter, before the December rollout completes.

RiskEarly warning in the buyer’s own recordsOwner
A name mistaken for a controlConnected agents reported as governed when Microsoft can only see them.A board or audit report that counts listed agents as controlled.CISO
A stale registryThe registry drifts from each platform between manual syncs.Agents deleted on a platform still listed in Agent 365.Head of identity and access management
Standing connector keysA stored key for each rival platform becomes a high-value secret.Connector keys with no vault entry or rotation date.Head of identity and access management
Agents beyond managed devicesDeveloper agents run where Defender and Intune do not reach.Agent use on devices missing from the managed inventory.Head of endpoint engineering
Liability no one has assignedA listed agent misbehaves and no contract says who must stop it.Platform contracts silent on revocation time and evidence.General counsel
Against this: This is a judgment about order, not a prediction of harm.
Source: Greyhound Research analytical model, not measured data. Greyhound Research analysis of the evidence in this research note.

Table 5 lists five risks in using one directory for agents across vendors, each with an early warning in the buyer’s own records and an owner. A name mistaken for a control, warned by a report counting listed agents as controlled, owned by the CISO. A stale registry, warned by deleted agents still listed, owned by the head of identity and access management. Standing connector keys, warned by keys with no vault entry or rotation date, owned by the same head of identity. Agents beyond managed devices, warned by agent use on unmanaged devices, owned by the head of endpoint engineering. Liability no one has assigned, warned by contracts silent on revocation, owned by the general counsel. The risks are Greyhound Research analysis.

Table 5. Every risk shows first as two lists that disagree. Current through November 19, 2026. © Greyhound Research.

The five share a cause: the belief that an agent listed in Agent 365 is an agent Microsoft governs. Each warning sits in a document the buyer already holds, such as a board report, a platform’s agent list, a key vault, a device inventory, or a contract. The table names owners, not likelihoods, and ranks none of the risks above another. All five can be checked before the connected platform rollout completes in December.

Greyhound Standpoint:

In Greyhound Research’s reading, none of the five risks is new. Each is an established identity problem, from stale accounts to standing keys, arriving at the scale and speed that agents bring. That is reassuring, because identity teams already know how to find these problems, and demanding, because the volume can outrun reviews built for human accounts. Greyhound CIO Pulse 2025 found 51% of enterprises already using multiple agents reporting task confusion, duplicated effort, and conflicting instructions, an early form of the same drift.

The common cause is the belief that a listed agent is a governed one. Earlier waves of service accounts and machine identities showed how that belief settles in. Accounts were created for one task, kept their rights after the task ended, and came to light during an audit or an incident. Agents can follow the same path faster, since they are made in minutes and may act with no person watching. The evidence supports the order of these risks, not their likelihood, and this note ranks none above another. Low drift in the first reconciliations, or a Microsoft enforcement matrix, would lighten the weight on several of them.

Risk owners should check all five before the connected platform rollout completes in December, and the CISO should take the results to the next board risk review, with each risk tied to its Runbook move.

↑ Back to the section guide

Part 5The Risks And The Counter-Argument

The Strongest Objection: One List Is Better Than None

Greyhound SpotlightThe strongest objection to this note is that most companies cannot list their agents at all, so a read-only list is a large step forward. That part stands.

The objection deserves its full weight. Every security program starts with knowing what exists, and almost no company can list its agents across platforms today. Microsoft already holds the employee directory, so it is the natural place to list the agents those employees sponsor. Seeing a Salesforce agent’s activity beside a Copilot agent’s is better than seeing neither. On this view, enforcement can follow later.

The objection has history on its side. Security programs have usually built visibility first and control second, and many of today’s enforcement tools began life as monitoring products. Greyhound Research anticipates that Microsoft will add enforcement where it owns the ground and where partners agree, as vendors tend to once a preview matures. The difficulty is the period in between, which may run for several release cycles. During it, the company’s own descriptions of its coverage, in audit responses, insurance forms, and customer security questionnaires, will need to separate observed agents from controlled ones, whatever the console shows. The question for buyers is less whether enforcement will follow than how they describe their coverage until it does.

Exhibit 6 tests each part of the objection against the evidence. Three parts stand. The fourth, that enforcement can follow later, falls on how directories are read.

Greyhound Research · Exhibit 6 · The objection held to the evidence

Three parts of the objection stand, and one falls

The fair counter-case to this research note, part by part, against the evidence gathered for each.

The objection holds for inventory, not for assurance. Use the list, and label connected agents as observed.

StandsFalls
Inventory is the first controlStandsNo program can govern agents it cannot list, and few companies can list them today.
Microsoft holds the workforce directoryStandsEntra already holds the people who sponsor each agent, so one list can link both.
Seeing beats blindnessStandsActivity from five platforms on one screen is better than none.
Enforcement can follow laterFalls on how lists are readA list shown to boards is read as control, and the documentation says it cannot add controls.
Source: Greyhound Research analytical model, not measured data. Greyhound Research analysis of Microsoft Learn and Microsoft’s Agent 365 post.

Exhibit 6 holds four parts of the strongest objection to this research note against the evidence, as four cards, and shows that three stand and one falls. Inventory is the first control: stands. Microsoft holds the workforce directory: stands. Seeing beats blindness, with activity from five platforms on one screen: stands. Enforcement can follow later: falls on how lists are read, because a list shown to boards is read as control and Microsoft’s documentation says it cannot add controls. The verdicts are Greyhound Research analysis.

Exhibit 6. Three parts of the objection stand, and one falls. Current through November 19, 2026. © Greyhound Research.

Greyhound Standpoint:

Greyhound Research accepts the objection for inventory and rejects it for assurance. Three of its four parts stand. Inventory is the first control, Microsoft holds the workforce directory, and seeing a rival’s agent beats not seeing it. Greyhound CIO Pulse 2025 found 49% of global CIOs looking to consolidate their AI experimentation pipelines onto a single governance platform, and Microsoft is answering that demand with the best placed list available.

The fourth part, that enforcement can follow later, is where the objection falls, and the reason lies less in the product than in how lists are read. A registry shown in audit and board reporting tends to be taken as proof of control. A board shown many platforms’ agents on one Microsoft screen will reasonably assume Microsoft can stop them. Enterprise technology has seen this sequence with consolidated dashboards before, where visibility arrived first and was read as control until an incident tested it. The objection would gain ground if Microsoft published which controls reach each connected platform, or if the console marked connected agents as observed. Both are within Microsoft’s gift.

Until then, the CISO and the audit committee should agree a reporting rule at the next board risk review: connected agents appear as observed, with the platform that can stop each one named beside them. One list is better than none, provided the screen says what it does.

The Deeper Read · About 1 minute

Select a question to open the full answer

1Why does the last part of the objection fall on how the list is read?

Lists become assurances. A registry shown in audit and board reporting tends to be read as proof of control. Microsoft’s own documentation says the connected view cannot add controls the source platform lacks. Greyhound Research found in May 2025 that 49% of global CIOs surveyed for Greyhound CIO Pulse 2025 wanted a single governance platform for AI experimentation, which explains the appeal of one screen and the risk of reading too much into it.

Accountability is the buyer’s frame. Greyhound Research’s November 5 note on agent front doors cited Greyhound CIO Pulse 2026, a survey of more than 1,000 enterprise technology leaders, that 84 percent have shifted their focus from AI adoption to AI accountability. Accountability needs a party that can stop the agent, not only one that can see it.

What the two survey findings tell a buyer. Read side by side, the two figures describe a tension many programs are likely to feel. Leaders want one place to govern AI, and they also want someone to answer for what AI does. A single list serves the first wish well. It serves the second only where the list’s owner can act. Buyers can hold both by asking one question of every governance screen: for each row, who can stop this agent, and how quickly?

In the source’s own words

  • Greyhound Research, Build 2025 software lifecycle note49% of global CIOs are now looking to consolidate their AI experimentation pipelines onto a single governance platformSource
  • Greyhound Research, November 5 front door note84 percent have shifted their focus from AI adoption to AI accountabilitySource

↑ Back to the section guide

Part 6What To Do, By Role

The Greyhound CXO Runbook: Ten Moves, None Waiting For Microsoft

Greyhound SpotlightEvery move uses a record the buyer already controls: an account register, an agent list, a gateway rule, a device policy, a contract, a count, a pilot approval, a registry, a device inventory, or a board report.

Ten moves follow from the evidence, at least one for each leader who will own part of a cross-vendor agent program. 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 deliberately ordinary. Each applies a discipline that identity and security teams already practice for people and service accounts, which is why none should need a new product to begin. Programs of this kind have tended to stall when they open with tool selection, and to drift when they open with policy papers. Starting with a register, an enforcement map, and a count gives the program something to measure from its first month. The ten are also built to outlast a change of vendor. If a rival’s registry later becomes the record, or Microsoft adds enforcement on connected platforms, the register, the enforcement map, and the contract terms carry over without being rewritten.

Table 6 sets the ten moves beside the assumption each corrects, its evidence, and its owner. Two of the ten finish only when Microsoft publishes more.

Greyhound Research · Table 6 · 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 a cross-vendor agent program, with the assumption each corrects, its evidence, and its owner.

Start all ten this quarter. Only the enforcement map and the contract wait on Microsoft to finish.

Starts on controls the buyer already holdsCompletes when Microsoft documents more

Select a move to open its full text in The Deeper Read below

MoveThe assumption it correctsEvidence anchorThe decision sits with
1/Register every agent as a privileged account with a named sponsorA listed agent is a governed agentExhibit 1CISO
2/Map where each agent’s credentials are issuedThe directory holds the kill switchExhibit 2Head of identity and access management
3/Keep tool-level limits on each runtime or in a company gatewayOne directory can replace each runtime’s rulesTable 4Chief enterprise architect
4/Contain local agents before naming themInspection alone stops a local agentExhibit 3Head of endpoint engineering, with the CISO
5/Write the enforcement duty into each contractA listed agent is someone else’s liabilityExhibit 4General counsel, with procurement
6/Count your own agents, and reconcile monthlyVendor counts measure coverageTables 2 and 3CIO
7/Approve the pilot on the documented method, not the announced oneThe announcement describes what engineers will configureTable 1CIO, with the head of identity and access management
8/Name one directory of record, and let the other registries feed itSeveral registries can each serve as the recordExhibit 5Chief enterprise architect
9/Find the agents that run beyond managed devicesDevice controls reach every machine an agent runs onTable 5Head of endpoint engineering
10/Require every coverage statement to separate observed agents from controlled onesThe console view can stand as a coverage statementExhibit 6Chair of the audit committee
Source: Greyhound Research analysis of the evidence in this research note.

Table 6 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 documents more. The moves are: register every agent as a privileged account with a named sponsor, map where each agent’s credentials are issued, keep tool-level limits on each runtime or in a company gateway, contain local agents before naming them, write the enforcement duty into each contract, count agents and reconcile monthly, approve the pilot on the documented method rather than the announced one, name one directory of record and let the other registries feed it, find the agents that run beyond managed devices, and require every coverage statement to separate observed agents from controlled ones. The moves are Greyhound Research analysis.

Table 6. Ten moves, and none waits on Microsoft to start. Current through November 19, 2026. © Greyhound Research.

The two that finish later, on the enforcement map and on the contract, need Microsoft to document what each connected platform enforces. Those answers 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 5 minutes

Select a move to open its full text

1CISOs: register every agent as a privileged account with a named sponsor.

Exhibit 1 shows that a directory entry is an inventory, not a control. This move corrects the belief that a listed agent is a governed one. Its action is to treat every agent, on every platform, as a privileged account: a named human sponsor, a written list of permitted tools, approval before any action that cannot be undone, and a quarterly review that removes unused rights. The model is privileged access, not employment: agents are approved and reviewed like privileged accounts, and are not priced, credited, or managed as staff. It agrees with the companion note on Azure agent infrastructure. The decision sits with the CISO.

2Identity leads: map where each agent’s credentials are issued.

Exhibit 2 shows that an agent can be stopped only where its pass is issued. This move corrects the assumption that the directory holds the kill switch. Its action is to record, for each agent, where it is listed and who issues its credentials, and to name in the incident plan the platform and person that can revoke it. The decision sits with the head of identity and access management.

Why the first two moves come first. The register and the enforcement map are the foundation for the other eight. Tool-level rules, device containment, contract terms, and monthly counts all depend on knowing which agents exist, who sponsors them, and who can revoke them. A company that writes contract clauses before that map exists will assign duties for an estate it has not yet described. Completing the first two moves this quarter gives every later move a list to work from.

3Enterprise architects: keep tool-level limits on each runtime or in a company gateway.

Table 4 shows AWS enforcing generally available rules on each tool call, Google documenting such rules without a single release label, and Microsoft’s equivalent in preview. This move corrects the belief that one directory can replace each runtime’s policy engine. Its action is to set tool-level rules where the agent runs, or in a gateway the company controls, and to use Agent 365 to list agents and sponsors. The decision sits with the chief enterprise architect.

4Endpoint leads: contain local agents before naming them.

Exhibit 3 shows Microsoft’s strongest controls on the device and a block that can arrive after a tool runs. This move corrects the belief that inspection alone stops a local agent. Its action is to set Execution Container boundaries through Intune, block unapproved agents first, run runtime protection in audit mode before block mode, and keep narrow rights for agents that can delete or send data. The decision sits with the head of endpoint engineering, together with the CISO.

5General counsel and procurement: write the enforcement duty into each contract.

Exhibit 4 shows Microsoft’s role written as read-only and the platform owner holding the credentials. This move corrects the assumption that a listed agent is someone else’s liability. Its action is to state in each platform’s contract who can stop an agent and within what time, and whose log is the record, and to record in the Agent 365 order that connected agents are observed only during preview. The decision sits with the general counsel, together with the head of procurement.

6CIOs: count your own agents, and reconcile monthly.

Table 2 and Table 3 show a headline count without a definition and a registry that syncs on request. This move corrects the habit of reading vendor counts as coverage. Its action is to count agents on every platform, compare the counts with Agent 365 every month, and report active and sponsored agents rather than registered ones. The decision sits with the CIO.

7CIOs: approve the connected platforms pilot on the documented method, not the announced one.

Table 1 shows most of the cross-vendor pieces in preview, and three items carrying different statuses on different Microsoft pages. This move corrects the assumption that the announcement describes what engineers will configure. Its action is to approve connected platforms as an inventory pilot on two or three of the seven documented platforms, and to plan, budget, and assess its risk on the documented method, which today is a stored key for every platform. Keyless sign-in stays a promised improvement until Microsoft confirms it in writing. The pilot budget should carry two costs that do not appear in Microsoft’s price: the other provider’s charges for the queries that read activity data, and the staff time to rotate each key. Frontier previews fall under the preview terms of the customer’s agreements, and the approval should say so. The decision sits with the CIO at the pilot gate, together with the head of identity and access management.

8Enterprise architects: name one directory of record, and let the other registries feed it.

Exhibit 5 shows AWS, ServiceNow, Okta, and Microsoft each bidding in 2026 to list agents beyond their own platforms, with AWS’s Agent Registry generally available since August 31 and Microsoft’s connected platforms in preview. This move corrects the assumption that several registries can each serve as the record. A company is unlikely to run five master lists well, and the cost of the overlap shows up as reconciliation work. Its action is to start from the registries the company already runs, name one of them as the directory of record for agents, and let the others feed it. Before testing anything else, architects should test how each candidate handles a sponsor who leaves, and favor the directory that can show, platform by platform, what it enforces as well as what it lists. The decision sits with the chief enterprise architect at the next architecture review, read against any news from AWS’s re:Invent, which opens November 30.

9Endpoint leads: find the agents that run beyond managed devices.

Table 5 names agents beyond managed devices as one of five risks, with an early warning in the buyer’s own records: agent use on devices missing from the managed inventory. This move corrects the assumption that device controls reach every machine an agent runs on. Defender, Intune, and Execution Containers act on devices the company manages, and even there coverage has two tiers. Defender discovers agents on managed Windows and macOS devices, while containment and runtime protection are documented for Windows. Its action is to compare the managed device inventory with every record of agent use the company already holds, treat each agent found on an unmanaged device as a finding with a named owner, and either bring that device under management or keep the agent away from company data. The decision sits with the head of endpoint engineering, who should take the results to the next board risk review with the other risks in Table 5.

10Audit committee chairs: require every coverage statement to separate observed agents from controlled ones.

Exhibit 6 shows three parts of the strongest objection standing and one falling, because a list shown to boards is read as control while Microsoft’s documentation says the connected view cannot add controls the source platform lacks. This move corrects the assumption that what the console shows can stand as a statement of coverage. Its action is to adopt one reporting rule at the next board risk review, under which connected agents appear as observed, with the platform that can stop each one named beside them. The same rule should govern every description of coverage the company gives outside the board, in audit responses, insurance forms, and customer security questionnaires. Each statement should answer, row by row, who can stop the agent and how quickly. The rule can be relaxed platform by platform as Microsoft documents controls that reach the source. The decision sits with the chair of the audit committee.

↑ Back to the section guide

Part 6What To Do, By Role

Questions For Microsoft: What Buyers Will Ask Next

Greyhound SpotlightFour of the six questions are answered in part by Microsoft’s own pages, which give different answers.

Every finding in this note rests on what Microsoft has published, and Microsoft alone can close the gaps it names. Six 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.

Each of the six can be put to an account team in writing, and the form of Microsoft’s answer will itself be informative. A documented answer on a Microsoft Learn page can be cited in a contract and tested by engineers. A spoken assurance in a briefing cannot. Greyhound Research expects the December rollout to settle the roster and status questions first, since the documentation has to catch up with the announcement for the rollout to complete. The questions on enforcement and responsibility are harder, because they ask Microsoft to define its role on platforms it does not run. Buyers who ask together, ahead of a renewal, are likely to receive fuller answers than one buyer asking alone.

Table 7 tracks the six questions against what Microsoft has said so far. Four are answered in part, and two not at all.

Greyhound Research · Table 7 · Questions for Microsoft, tracked

Four questions answered in part, two not at all

The six questions only Microsoft can answer, against what it had stated by 00:40 UTC on November 19, 2026.

Put all six to the account team in writing. The per-platform enforcement answer matters most, and no move waits on it.

Stated in partNot stated

Select a question to open its full text in The Deeper Read below

QuestionWhat Microsoft has said so farThe Runbook move it settles
Q1What does the 50 million count?Stated in part Registered since May 2026. Active, unique, and source not stated.Runbook move 6
Q2For each connected platform, which controls are enforced and which only observed?Stated in part Read-only, and cannot add controls. No per-platform list.Runbook move 2
Q3Does a connected agent sign in with a stored key or federation, and does its Agent ID carry Conditional Access?Stated in part Two answers: keyless in the announcement, stored key in the documentation.Runbook move 2
Q4When will OpenAI, SAP, ServiceNow, and UiPath be documented, with scheduled sync and deletion?Not stated Named at Ignite, absent from the documentation.Runbook move 6
Q5Is runtime protection generally available, and when will a block stop a tool before it runs?Stated in part Two statuses, and a documented post-tool gap.Runbook move 4
Q6Who is responsible when a connected agent breaks a policy Agent 365 displays?Not stated Platform administrators own credentials. No statement on third-party terms.Runbook move 5
Against this: None of the six is a trick question. Each asks Microsoft to write down what its product does or will do at release.
Source: Greyhound Research analysis of Microsoft’s Book of News, Agent 365 post, and Microsoft Learn.

Table 7 tracks six questions only Microsoft can answer against what it had stated by 00:40 UTC on November 19, 2026, and shows four answered in part and two not stated. Answered in part: what the 50 million counts, which controls are enforced on each connected platform, how connected agents sign in, and whether runtime protection is generally available. Not stated: when OpenAI, SAP, ServiceNow, and UiPath will be documented with scheduled sync and deletion, and who is responsible when a connected agent breaks a policy Agent 365 displays. Each question names the Runbook move it settles. The grading is Greyhound Research analysis.

Table 7. Four questions answered in part, two not at all. Current through November 19, 2026. © Greyhound Research.

Two answers would change this note’s reading most. One is a published table, platform by platform, of which controls Agent 365 and Entra enforce on connected agents and which they only observe. The other is a definition of the 50 million, split by active and registered agents and by source. Either would move Microsoft from the broadest directory to the most trusted one. None of the six is a trick question. Each asks Microsoft to write down what its product does or will do at release.

The Deeper Read · About 2 minutes

Select a question to open its full text

Q1What does the 50 million count?

Said so far. In part. Microsoft says the agents were registered since its May 2026 launch. It has not said how many are active, unique, or from other vendors. The answer settles. Whether the figure measures governance or creation, under the sixth Runbook move. Evidence that would settle it. A count split by active and registered agents, and by source platform. Greyhound Research would update. Table 2.

Q2For each connected platform, which controls are enforced and which are only observed?

Said so far. In part. The documentation says the connection is read-only and cannot add capabilities the platform lacks. It does not give a per-platform list. The answer settles. Where the enforcement map of the second Runbook move can show Microsoft as able to stop an agent. Evidence that would settle it. A published enforcement matrix by platform. Greyhound Research would update. Exhibit 1.

Why the enforcement question comes first. Of the six, the second is the one whose answer reaches every other move. A published matrix would say, platform by platform, where Microsoft can act and where it can only watch. That single document would settle how boards are briefed, how incident plans are written, and how far contracts must go. Its absence is understandable while the feature is in preview, and it is still the gap buyers should press on first.

Q3Does a connected agent sign in with a stored key or with federation, and does its Entra Agent ID carry Conditional Access anywhere?

Said so far. In part, with two answers. The announcement says secretless federation. The documentation, updated November 6, says a stored key for every platform. The answer settles. Which secrets the identity team must vault and rotate, under the second Runbook move. Evidence that would settle it. Updated documentation per platform. Greyhound Research would update. Exhibit 2.

Q4When will OpenAI, SAP, ServiceNow, and UiPath be documented, and will sync run on a schedule and remove deleted agents?

Said so far. Not stated. The announcement names all four, and the documentation lists none of them. The answer settles. How often the reconciliation of the sixth Runbook move must run. Evidence that would settle it. Documentation for each platform and a stated sync and deletion policy. Greyhound Research would update. Table 3.

Q5Is runtime protection for local agents generally available or in preview, and when will a block stop a tool before it runs?

Said so far. In part, with two answers. The Book of News and the Agent 365 post say generally available on Windows. The Microsoft Learn page says preview and records a post-tool block gap for Claude Code. The answer settles. Whether the fourth Runbook move can switch from audit to block. Evidence that would settle it. An updated support matrix and release status. Greyhound Research would update. Exhibit 3.

Q6Who is responsible when a connected agent breaks a policy that Agent 365 displays?

Said so far. Not stated. The documentation gives the platform administrator ownership of credentials and makes no statement on third-party terms. The answer settles. What the contract of the fifth Runbook move must assign. Evidence that would settle it. A written statement of Microsoft’s role in Agent 365’s terms. Greyhound Research would update. Exhibit 4.

↑ Back to the section guide

Part 7The Final Greyhound Standpoint

Agent Identity: The Final Greyhound Standpoint

Greyhound SpotlightFour signals within twelve months would move this position, and Microsoft controls all four.

Greyhound Research’s position is that Ignite 2026 made Microsoft’s Entra the strongest candidate for the directory of every company’s agents, and that the directory, however complete, does not by itself decide what those agents may do. Rivals’ agents now arrive with a Microsoft name and a human sponsor. The name does not bring the power to stop them. When rivals’ agents get a Microsoft name, enforcement becomes the control, and enforcement lives wherever each agent’s credentials are issued.

On reach, Microsoft leads. It names eleven rival platforms, finds more than sixty kinds of local agent, and ties every agent to a person in the directory that already holds the workforce. On enforcement at the tool call, AWS leads today with generally available rules, as the companion note on Azure agent infrastructure found, and Google documents tool-level rules without a single release label. On rivals’ platforms, no vendor documents enforcement, and Microsoft’s own documentation says it does not act there. On the device, Microsoft enforces more than any other vendor read for this note, with a gap it records itself.

The shape of this contest is familiar. Platform shifts in enterprise technology have repeatedly produced a contest over the system of record, and the usual outcome has been one directory that others feed, with control kept close to where the work runs. Agents look set to follow that course. What differs is speed. Agents are created in minutes and act without a person watching, so the gap between a list and a control matters sooner than it did for accounts or devices.

Microsoft deserves credit on three counts. It built the cross-vendor list openly against its competitors’ products, it documented the limits of that list in plain terms, and it shipped its strongest controls where it owns the ground, on the managed device. The evidence still has limits. Most of the items are previews. Microsoft’s announcement and its documentation disagree on how connected platforms sign in, on which platforms are supported, and on the status of runtime protection. The documentation may catch up before the December completion date. No contract or terms of service were read. Statuses and claims are as each company published them through November 19, 2026.

Four signals would move this position within twelve months. Each would strengthen the case for using Agent 365 as more than an inventory. They are a per-platform enforcement matrix, scheduled sync with deletion and keyless sign-in, a defined count of active agents, and runtime protection that blocks before a tool runs.

Exhibit 7 sets those signals side by side with what each would change. All four sit with Microsoft.

Greyhound Research · Exhibit 7 · Signals that would move the position

Four signals in twelve months would move this position

What would move Greyhound Research’s judgment within twelve months, and what each would change.

Of the four signals, a per-platform enforcement matrix would move this view furthest. Watch for it first.

Signal Microsoft controlsEvent outside its control
A per-platform enforcement matrixA published table of which controls Agent 365 and Entra enforce on each connected platform.Would settle whether the name comes with a way to stop the agent. Controlled by: Microsoft.A registry that matches realityScheduled sync, removal of deleted agents, and keyless sign-in for every platform.Would let the registry stand as the record without a manual check. Controlled by: Microsoft.A defined countThe 50 million split by active and registered agents and by source platform.Would turn a creation figure into a governance figure. Controlled by: Microsoft.A block before the tool runsRuntime protection that stops every supported local agent before a tool acts.Would let endpoint teams move from audit to block. Controlled by: Microsoft.Outside Microsoft’s controlNews on agent identity or registries at AWS’s re:Invent, which opens November 30, could reset the comparison.
Source: Greyhound Research analytical model, not measured data. Greyhound Research analysis.

Exhibit 7 lists the four signals that would move Greyhound Research’s position within twelve months, each controlled by Microsoft, and one event outside its control. A per-platform enforcement matrix would settle whether the name comes with a way to stop the agent. A registry with scheduled sync, removal of deleted agents, and keyless sign-in would let it stand as the record. A defined count would turn the 50 million into a governance figure. Runtime protection that blocks before a tool runs would let endpoint teams move from audit to block. News on agent identity or registries at AWS’s re:Invent, from November 30, could reset the comparison. The signals are Greyhound Research analysis.

Exhibit 7. Four signals in twelve months would move this position. Current through November 19, 2026. © Greyhound Research.

One event outside Microsoft’s control could reset the comparison: news on agent identity or registries at AWS’s re:Invent, which opens November 30. Of the four signals, a per-platform enforcement matrix would matter most, because it decides whether the name in the directory comes with a way to stop the agent.

The decisions fall to named owners at known points. CISOs should take the rule that connected agents are reported as observed, not governed, to the next board risk review. Identity leads should finish the enforcement map and start the monthly reconciliation before the December rollout completes. Enterprise architects should settle the directory of record and the home of tool-level rules at the next architecture review, and general counsel should carry the enforcement duty into each platform’s next renewal. Until the signals arrive, companies should list every agent in one place, and keep the power to stop each one where its credentials are issued.

↑ Back to the section guide

Part 8Appendix, 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. Every source was read on November 18, 2026, and re-verified on November 19, 2026, between 00:20 and 00:40 UTC, and the sources added on November 19 were read in the same window. 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 blogs7 sources, 41 extracts
Source, read, and usedWhere in the sourceWhat it says
Althoff, J. (2026)Read November 18, 2026. Re-verified November 19, 2026, between 00:20 and 00:40 UTCUsed in What Microsoft Announced, The 50 Million. Table 1. Exhibit 5Quoted in The Deeper Read of What Microsoft Announced, The 50 MillionPost[quoted in The Deeper Read]
Post"General availability of the new Microsoft 365 E7: The Frontier Suite on May 1 for $99 per user"
PostAgent 365 is described as the control plane for AI agents (summarized, not quoted).
Preview customers[quoted in The Deeper Read]
Davuluri, P. (2026)Read November 18, 2026. Re-verified November 19, 2026, between 00:20 and 00:40 UTCUsed in Local And Shadow Agents. Table 1. Exhibit 3Quoted in The Deeper Read of Local And Shadow AgentsPost[quoted in The Deeper Read]
PostProcess container improvements are in preview, learning and permissive modes are generally available, and Local Micro VMs are experimental (summarized, not quoted).
Reading noteThe post names Anthropic among builders on MXC and does not mention Claude Code or Defender runtime protection by name.
Holeček, A. (2026)Read November 18, 2026. Re-verified November 19, 2026, between 00:20 and 00:40 UTCUsed in A Name Is Not A Control, Where Enforcement Lives. Table 1Quoted in The Deeper Read of A Name Is Not A Control, Where Enforcement LivesSecure agentic insiders[quoted in The Deeper Read]
Same section[quoted in The Deeper Read]
Same section[quoted in The Deeper Read]
Same section"These capabilities remain in limited preview."
DiscoveryDiscovery expands by 35 local agent types to 60 in total, including Claude Code, OpenAI Codex, and GitHub Copilot, plus third-party cloud agents on AWS and Google Cloud (summarized, not quoted).
Microsoft (2026a)Read November 18, 2026. Re-verified November 19, 2026, between 00:20 and 00:40 UTCUsed in What Microsoft Announced, Local And Shadow Agents, Three Clouds Compared. Tables 1 and 4. Exhibit 3Quoted in The Deeper Read of What Microsoft Announced, Local And Shadow AgentsMicrosoft Entra[quoted in The Deeper Read]
Microsoft Entra[quoted in The Deeper Read]
Microsoft Entra"will coordinate Intune, Defender and Global Secure Access policies through Agent 365."
Microsoft EntraAgentic traffic acquisition is generally available, and Agent ID for Microsoft cloud agents, extended to Autopilot, is in preview (summarized, not quoted).
Intune[quoted in The Deeper Read]
Intune[quoted in The Deeper Read]
Defender"Agent-native runtime protection on Windows is generally available for Claude Code, GitHub Copilot CLI,"
Defender[quoted in The Deeper Read]
Agent 365 governance[quoted in The Deeper Read]
Management insights"An Agent 365 license is still required for customers to get enterprise-grade management, security and governance."
Reading noteThe Book of News does not mention connected platforms or the figure of 50 million.
Microsoft (2026b)Read November 18, 2026. Re-verified November 19, 2026, between 00:20 and 00:40 UTCUsed in The 50 Million. Table 2Quoted in The Deeper Read of The 50 MillionPost[quoted in The Deeper Read]
Post"It is a control plane that extends your existing governance, identity, security and management systems"
Reading noteThe post is an edited version of Satya Nadella's Ignite keynote remarks, with the byline Microsoft Corporate Blogs. It does not mention Entra, connected platforms, AWS, Google, Anthropic, OpenAI, SAP, or ServiceNow.
Microsoft (2026h)Read November 19, 2026, between 00:20 and 00:40 UTCUsed in The 50 MillionQuoted in The Deeper Read of The 50 MillionStoryAs of September 2026, Insight had discovered 9,734 agents across its environment (summarized, not quoted).
Story[quoted in The Deeper Read]
Reading noteDated October 29, 2026. The quotation is by Will Pocknell, Director, Cybersecurity, Insight Enterprises. The story does not name the tool that discovered the agents.
Stanford, C. (2026)Read November 18, 2026. Re-verified November 19, 2026, between 00:20 and 00:40 UTCUsed in What Microsoft Announced, The 50 Million, A Name Is Not A Control, The Connected Roster, Local And Shadow Agents. Tables 1, 2, and 3. Exhibits 1, 2, and 3Quoted in The Deeper Read of What Microsoft Announced, The Connected Roster, The 50 Million, Local And Shadow AgentsConnected platforms[quoted in The Deeper Read]
Connected platforms[quoted in The Deeper Read]
Connected platforms[quoted in The Deeper Read]
Connected platformsThe rollout extends to Anthropic Claude Managed Agents, OpenAI, Oracle, SAP, ServiceNow, and UiPath and will complete in December, marked preview with capabilities that vary by platform (summarized, not quoted).
Scale[quoted in The Deeper Read]
Local agents[quoted in The Deeper Read]
Local agents[quoted in The Deeper Read]
Local agents"Local agent discovery in Microsoft Defender is now generally available"
Runtime protection"Now generally available on Windows, Defender inspects critical points in the agent loop"
Action center[quoted in The Deeper Read]
Runtime protection[quoted in The Deeper Read]
Reading noteThe post does not say whether controls on connected agents are enforced on the third-party platform. Defender posture assessment for Bedrock, Gemini, Vertex, and Claude agents is described as risk assessment and recommendations, in preview.
Microsoft documentation5 sources, 26 extracts
Source, read, and usedWhere in the sourceWhat it says
Microsoft (2026c)Read November 18, 2026. Re-verified November 19, 2026, between 00:20 and 00:40 UTCUsed in What Microsoft Announced, A Name Is Not A Control, The Connected Roster, When Enforcement Fails, The Strongest Objection. Tables 1 and 3. Exhibits 1, 2, and 4Quoted in The Deeper Read of A Name Is Not A Control, What Microsoft Announced, The Connected RosterIntroduction[quoted in The Deeper Read]
Security model[quoted in The Deeper Read]
Security model[quoted in The Deeper Read]
Authentication[quoted in The Deeper Read]
Limitations[quoted in The Deeper Read]
Limitations[quoted in The Deeper Read]
Terms[quoted in The Deeper Read]
Supported platformsSeven platforms are listed: Amazon Bedrock, Google Vertex AI, Salesforce Agentforce, Databricks Genie, Anthropic Claude Managed Agents, Oracle Generative AI Agents, and Snowflake Cortex, with activity data coming soon for Oracle and Snowflake (summarized, not quoted).
ResponsibilitiesThe platform administrator owns creating credentials, granting permissions, enabling telemetry, and rotating or revoking credentials (summarized, not quoted).
Reading noteLast updated November 6, 2026. The page does not describe an Entra Agent ID being created for synchronized agents, and makes no statement on third-party terms.
Microsoft (2026d)Read November 18, 2026. Re-verified November 19, 2026, between 00:20 and 00:40 UTCUsed in What Microsoft Announced, Where Enforcement Lives. Tables 1 and 4Quoted in The Deeper Read of What Microsoft Announced, Where Enforcement LivesLicensing"Agent ID is available for all Microsoft Entra customers."
Licensing[quoted in The Deeper Read]
Third-party agents[quoted in The Deeper Read]
Reading noteThe page shows a last-updated date of June 24, 2026, and gives no release label.
Microsoft (2026e)Read November 18, 2026. Re-verified November 19, 2026, between 00:20 and 00:40 UTCUsed in Where Enforcement Lives, Every Vendor Wants The DirectoryQuoted in The Deeper Read of Where Enforcement LivesSponsors[quoted in The Deeper Read]
Sponsors[quoted in The Deeper Read]
LicensingGovernance of agent identities requires Microsoft 365 E7, or Agent 365 with at least Entra P1 or Microsoft 365 E3 (summarized, not quoted).
Microsoft (2026f)Read November 18, 2026. Re-verified November 19, 2026, between 00:20 and 00:40 UTCUsed in Where Enforcement Lives. Exhibit 2Quoted in The Deeper Read of Where Enforcement LivesPlatform paths[quoted in The Deeper Read]
Federation[quoted in The Deeper Read]
Federation[quoted in The Deeper Read]
Reading noteLast updated November 17, 2026. The page does not mention Conditional Access or say whether Entra controls apply to systems outside Microsoft.
Microsoft (2026g)Read November 18, 2026. Re-verified November 19, 2026, between 00:20 and 00:40 UTCUsed in What Microsoft Announced, Local And Shadow Agents. Table 1. Exhibit 3Quoted in The Deeper Read of What Microsoft Announced, Local And Shadow AgentsNotice[quoted in The Deeper Read]
Modes[quoted in The Deeper Read]
Known issues[quoted in The Deeper Read]
Known issuesPost-tool-response inspection occurs after the tool runs and cannot undo actions the tool has already completed (summarized, not quoted).
Reading noteThe page is titled preview and was last updated November 11, 2026. Network inspection does not support agents that use certificate pinning or HTTP/3.
AWS and Google9 sources, 15 extracts
Source, read, and usedWhere in the sourceWhat it says
Amazon Web Services (2025)Read November 18, 2026. Re-verified November 19, 2026, between 00:20 and 00:40 UTCUsed in Three Clouds Compared. Table 4AnnouncementAmazon Bedrock AgentCore is generally available, posted October 13, 2025, with AgentCore Identity named among its services (summarized, not quoted).
Amazon Web Services (2026a)Read November 18, 2026. Re-verified November 19, 2026, between 00:20 and 00:40 UTCUsed in Three Clouds Compared, Agent Identity. Table 4. Exhibit 5Quoted in The Deeper Read of Three Clouds ComparedAnnouncement[quoted in The Deeper Read]
AnnouncementPolicy in AgentCore is generally available, posted March 3, 2026, with rules converted to Cedar (summarized, not quoted).
Amazon Web Services (2026b)Read November 19, 2026, between 00:20 and 00:40 UTCUsed in Every Vendor Wants The Directory. Exhibit 5AnnouncementAWS Agent Registry, available through Amazon Bedrock AgentCore, entered preview on April 9, 2026, in five AWS Regions (summarized, not quoted).
Amazon Web Services (2026c)Read November 19, 2026, between 00:20 and 00:40 UTCUsed in Every Vendor Wants The Directory. Table 4. Exhibit 5AnnouncementAWS Agent Registry became generally available on August 31, 2026, as a private, governed catalog for agents, tools, skills, MCP servers, and custom resources within an organization (summarized, not quoted).
Reading noteThe announcement does not mention agents on premises or on other clouds. The AgentCore overview page does.
Amazon Web Services (n.d.)Read November 18, 2026. Re-verified November 19, 2026, between 00:20 and 00:40 UTCUsed in Every Vendor Wants The Directory. Table 4. Exhibit 5Quoted in The Deeper Read of Every Vendor Wants The DirectoryCore services[quoted in The Deeper Read]
Core servicesIdentity is described as an agent identity, access, and authentication management service compatible with existing identity providers (summarized, not quoted).
Sharma, R., Yuan, F., Khurpa, S. and Rodriguez, A. (2025)Read November 18, 2026. Re-verified November 19, 2026, between 00:20 and 00:40 UTCUsed in Every Vendor Wants The Directory. Table 4Quoted in The Deeper Read of Every Vendor Wants The DirectoryPost[quoted in The Deeper Read]
Google Cloud (2026a)Read November 18, 2026. Re-verified November 19, 2026, between 00:20 and 00:40 UTCUsed in Three Clouds Compared. Table 4Quoted in The Deeper Read of Three Clouds ComparedOverview[quoted in The Deeper Read]
Reading noteLast updated November 17, 2026. No release label on the page. Supported runtimes are all on Google Cloud.
Google Cloud (2026b)Read November 18, 2026. Re-verified November 19, 2026, between 00:20 and 00:40 UTCUsed in Every Vendor Wants The Directory. Table 4. Exhibit 5OverviewAgent Registry is described as the central hub for governance, discovery, and search for an AI ecosystem, with cross-cloud registration of agents on Kubernetes outside Google Cloud in limited preview (summarized, not quoted).
Google Cloud (2026c)Read November 18, 2026. Re-verified November 19, 2026, between 00:20 and 00:40 UTCUsed in Three Clouds Compared. Table 4Quoted in The Deeper Read of Three Clouds ComparedAccess[quoted in The Deeper Read]
MCP trafficThe gateway reads tool names and method attributes from MCP requests, enabling tool-level access policies (summarized, not quoted).
Reading noteLast updated November 18, 2026. No single release label. Cloud Run and cross-cloud use are in limited preview.
Okta, ServiceNow, and Zenity3 sources, 7 extracts
Source, read, and usedWhere in the sourceWhat it says
Okta (2026)Read November 18, 2026. Re-verified November 19, 2026, between 00:20 and 00:40 UTCUsed in Every Vendor Wants The Directory. Exhibit 5Quoted in The Deeper Read of Every Vendor Wants The DirectoryRelease[quoted in The Deeper Read]
ReleaseKnown agents can be imported from platforms like AWS Bedrock or Salesforce Agentforce. Agent SSO is generally available, and endpoint discovery of shadow agents is planned (summarized, not quoted).
ServiceNow (2026)Read November 18, 2026. Re-verified November 19, 2026, between 00:20 and 00:40 UTCUsed in Every Vendor Wants The Directory. Exhibit 5Quoted in The Deeper Read of Every Vendor Wants The DirectoryRelease[quoted in The Deeper Read]
Reading noteServiceNow's AI Control Tower release notes, read November 19, 2026, list the AWS, Microsoft, and Google connectors under early availability. General availability of discovery beyond ServiceNow was not confirmed.
ReleaseAI Control Tower discovers AI assets including systems beyond ServiceNow, with discovery integrations naming AWS, Google Cloud, and Microsoft Azure (summarized, not quoted).
Zenity (2026)Read November 19, 2026, between 00:20 and 00:40 UTCUsed in A Name Is Not A ControlQuoted in The Deeper Read of A Name Is Not A ControlMicrosoft statement[quoted in The Deeper Read]
Reading noteReleased November 17, 2026. The statement is by Ray Smith, Corporate Vice President, Agent 365 Product, Microsoft, and continues "into informed management, governance and security decisions". The release says Zenity provides telemetry and limited inline prevention for Copilot Studio agents on the Microsoft Copilot harness.
Media1 sources, 2 extracts
Source, read, and usedWhere in the sourceWhat it says
Constantin, A. M. (2026)Read November 18, 2026. Re-verified November 19, 2026, between 00:20 and 00:40 UTCUsed in The 50 Million. Table 2. Exhibit 5Quoted in The Deeper Read of The 50 MillionNadella post[quoted in The Deeper Read]
ArticleNadella’s posts, made after the earnings call and dated October 28, are embedded in the article, whose own summary says the agents “run in” Agent 365. None of them says whether the count is of registered or active agents (summarized, not quoted).
Research this note builds on8 sources, 13 extracts
Source, read, and usedWhere in the sourceWhat it says
Greyhound Research (2025a)Read November 18, 2026. Re-verified November 19, 2026, before 00:40 UTC, on greyhoundresearch.comUsed in The Strongest ObjectionQuoted in The Deeper Read of The Strongest ObjectionGreyhound CIO Pulse 2025[quoted in The Deeper Read]
Greyhound Research (2025b)Read November 18, 2026. Re-verified November 19, 2026, before 00:40 UTC, on greyhoundresearch.comUsed in The 50 Million, Every Vendor Wants The DirectoryQuoted in The Deeper Read of Every Vendor Wants The Directory, The 50 MillionIdentity section[quoted in The Deeper Read]
Greyhound CIO Pulse 2025[quoted in The Deeper Read]
NoteThe concern is named as silent privilege creep (summarized, not quoted).
Greyhound Research (2025c)Read November 18, 2026. Re-verified November 19, 2026, before 00:40 UTC, on greyhoundresearch.comUsed in Local And Shadow Agents, Risks To WatchQuoted in The Deeper Read of Local And Shadow AgentsGreyhound CIO Pulse 2025[quoted in The Deeper Read]
Greyhound CIO Pulse 2025"51% of enterprises already using multiple agents report problems with task confusion, duplicated effort, and conflicting instructions"
Note67 percent of IT leaders saw agent execution at the edge, on mobile, desktop, or browser, as the greatest blind spot in their AI risk posture (summarized, not quoted).
Greyhound Research (2026a)Read November 18, 2026. Re-verified November 19, 2026, before 00:40 UTC, on greyhoundresearch.comUsed in When Enforcement FailsQuoted in The Deeper Read of When Enforcement FailsClosing test[quoted in The Deeper Read]
Greyhound Research (2026b)Read November 18, 2026. Re-verified November 19, 2026, before 00:40 UTC, on greyhoundresearch.comUsed in The Strongest ObjectionQuoted in The Deeper Read of The Strongest ObjectionGreyhound CIO Pulse 2026[quoted in The Deeper Read]
NoteThe base is more than 1,000 enterprise technology leaders (summarized, not quoted).
Greyhound Research (2026c)Read November 18, 2026. Re-verified November 19, 2026, before 00:40 UTC, on greyhoundresearch.comUsed in When Enforcement FailsNoteConsolidation on one suite remains right while its standard terms fall short (summarized, not quoted). Listed as research this note builds on.
Greyhound Research (2026d)Draft read November 18, 2026. Not yet publishedUsed in A Name Is Not A Control, Where Enforcement Lives, Three Clouds Compared, The Greyhound CXO RunbookDraft, Agent Grid And IdentityIdentity is where Microsoft's record runs longest, AWS documents rules at each tool request, and every agent should be treated as a privileged account from its first day (summarized, not quoted).
Greyhound Research (2026e)Draft read November 18, 2026. Not yet publishedUsed in A Name Is Not A ControlDraftThe identity layer, not the confidential computing seal, is where an agent's rights are set (summarized, not quoted).

↑ Back to the section guide

Part 8Appendix, 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, Google’s, Okta’s, ServiceNow’s, and Zenity’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 news reports what its authors were told. Standing pages that carry no publication date are marked n.d. All sources were read, or re-read, on November 19, 2026.

Microsoft announcements and blogs

Althoff, J. (2026) Introducing the First Frontier Suite built on Intelligence + Trust. The Official Microsoft Blog, March 9. Available at: https://blogs.microsoft.com/blog/2026/03/09/introducing-the-first-frontier-suite-built-on-intelligence-trust/

Davuluri, P. (2026) Microsoft Ignite 2026: Building Windows for the next era of work. Windows Experience Blog, November 17. Available at: https://blogs.windows.com/windowsexperience/2026/11/17/microsoft-ignite-2026-building-windows-for-the-next-era-of-work/

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/

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) 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/

Microsoft (2026h) Insight Enterprises puts Security Copilot to work on 1,700-plus monthly triages. Microsoft Customer Stories, October 29. Available at: https://www.microsoft.com/en/customers/story/27530-insight-enterprises-microsoft-security-copilot

Stanford, C. (2026) What’s new in Microsoft Agent 365: Ignite 2026. Microsoft Tech Community, November 17. Available at: https://techcommunity.microsoft.com/blog/agent-365-blog/whats-new-in-microsoft-agent-365-ignite-2026/4571753

Microsoft documentation

Microsoft (2026c) Connected platforms in Microsoft Agent 365. Microsoft Learn, updated November 6. Available at: https://learn.microsoft.com/microsoft-agent-365/admin/connected-platforms

Microsoft (2026d) What is Microsoft Entra Agent ID? Microsoft Learn, updated June 24. Available at: https://learn.microsoft.com/en-in/Entra/agent-id/what-is-microsoft-entra-agent-id

Microsoft (2026e) Governing Agent Identities. Microsoft Learn, updated November 11. Available at: https://learn.microsoft.com/en-us/entra/id-governance/agent-id-governance-overview

Microsoft (2026f) Choose an integration path for an agent. Microsoft Learn, updated November 17. Available at: https://learn.microsoft.com/en-in/Entra/agent-id/configure-third-party-agents

Microsoft (2026g) AI agent runtime protection with Microsoft Defender for Endpoint (Preview). Microsoft Learn, updated November 11. Available at: https://learn.microsoft.com/en-us/defender-endpoint/ai-agent-runtime-protection-overview

AWS and Google

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 Agent Registry for centralized agent discovery and governance is now available in Preview. AWS What’s New, April 9. Available at: https://aws.amazon.com/about-aws/whats-new/2026/04/aws-agent-registry-in-agentcore-preview

Amazon Web Services (2026c) AWS Agent Registry for centralized agent discovery and governance is now generally available. AWS What’s New, August 31. Available at: https://aws.amazon.com/about-aws/whats-new/2026/08/aws-agent-registry-generally-available

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

Sharma, R., Yuan, F., Khurpa, S. and Rodriguez, A. (2025) Introducing Amazon Bedrock AgentCore Identity: Securing agentic AI at scale. AWS Machine Learning Blog, August 15. Available at: https://aws.amazon.com/blogs/machine-learning/introducing-amazon-bedrock-agentcore-identity-securing-agentic-ai-at-scale

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) Agent Registry. Google Cloud Documentation, updated November 17. Available at: https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/agent-registry

Google Cloud (2026c) 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

Okta, ServiceNow, and Zenity

Okta (2026) New Okta for AI Agents innovations increase visibility into agent behavior, secure connections at runtime, and enforce continuous agent governance. Okta Newsroom, September 22. Available at: https://www.okta.com/newsroom/press-releases/ai-innovations-oktane-2026/

ServiceNow (2026) ServiceNow expands AI Control Tower to discover, observe, govern, secure, and measure AI deployed across any system in the enterprise. Business Wire, May 5. Available at: https://www.businesswire.com/news/home/20260505712561/en/

Zenity (2026) Zenity Expands AI Agent Security Coverage Across the Microsoft Ecosystem. Business Wire, November 17. Available at: https://www.financialcontent.com/article/bizwire-2026-11-17-zenity-expands-ai-agent-security-coverage-across-the-microsoft-ecosystem

Media

Constantin, A. M. (2026) Microsoft’s Copilot daily users grew five times in a year, Nadella says. The Next Web, October 29. Available at: https://thenextweb.com/news/copilot-daily-users-nadella-foundry-agent-365

Research this note builds on

Greyhound Research (2025a) Microsoft Build 2025: Reimagining The Software Development Lifecycle With AI. Research note, May 21. Available at: https://greyhoundresearch.com/microsoft-build-2025-reimagining-the-software-development-lifecycle-with-ai/

Greyhound Research (2025b) 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 (2025c) 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 (2026a) Everyone Is Building AI Agents. Oracle Is Moving The Contest Inside The Runtime. Research note, July 20. Available at: https://greyhoundresearch.com/everyone-is-building-ai-agents-oracle-is-moving-the-contest-inside-the-runtime/

Greyhound Research (2026b) The Front Door Race – Oracle, OpenAI And Five Rivals, And Why The Permission Becomes The Control. Research note, November 5. Available at: https://greyhoundresearch.com/oracle-openai-the-front-door-race-the-permission-becomes-the-control/

Greyhound Research (2026c) 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/

Greyhound Research (2026d) Microsoft Azure Agent Infrastructure – When Agents Get Their Own Machines, Authority Becomes The Control. Research note, November 2026. Available at: https://greyhoundresearch.com/microsoft-azure-agent-infrastructure-when-agents-get-their-own-machines-authority-becomes-the-control/

Greyhound Research (2026e) Microsoft Confidential Computing – When Agents Work Inside The Seal, Attestation Becomes The Control. Research note, November 2026. Available at: https://greyhoundresearch.com/microsoft-confidential-computing-when-agents-work-inside-the-seal-attestation-becomes-the-control/

↑ Back to the section guide

Part 8Appendix, Sources, And Disclosures

The evidence in this research note is current through November 19, 2026, with an evidence cutoff of 00:40 UTC that day, when every source was re-verified, two days after Microsoft’s Ignite announcements. 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 security officers, heads of identity and access management, enterprise architects, chief information officers, heads of endpoint engineering, general counsel and procurement leaders, board members, and the press. It reads Microsoft’s agent identity and governance announcements at Microsoft Ignite 2026, the documentation beside them, and the published documentation of AWS, Google, Okta, and ServiceNow, to inform decisions on agent inventory, identity, enforcement, endpoint controls, and contracts. 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 liability 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, 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 thirty-three documents listed in Sources And References, in six 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 or not documented is bounded to the documents the text names, as they stood at 00:40 UTC on November 19, 2026. Statuses are given as each company states them, and where two Microsoft pages give different statuses, both are reported.

The figure of 50 million agents is Microsoft’s claim and is not used in the analysis. The Greyhound CIO Pulse figures are cited as published in Greyhound Research’s research notes of May 21, May 22, and May 26, 2025, and November 5, 2026, with the edition year stated. No composite figure is computed. Two earlier published views of Greyhound Research, from February and March 2026, are attributed to Greyhound Research in the text and are not listed as sources. No Greyhound Fieldnotes are quoted, and no factual claim in this research note rests on unpublished material. No contract or terms of service were read for this note beyond what Microsoft’s documentation states about preview terms.

The reading of what connected platforms enforce, where an agent’s credentials are issued, the duties in the liability exhibit, the readings in Exhibits 1 to 7 and Tables 1 to 7, 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.

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.

↑ Back to the section guide

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.

Discover more from Greyhound Research

Subscribe now to keep reading and get access to the full archive.

Continue reading