It's a big update, and the release states are uneven. Some pieces are generally available today. Many are in preview, and a few won't arrive until December. Admins planning around this need to keep those apart.
The scale argument: 50 million agents and counting
Microsoft frames the update with one number. Microsoft says organizations have registered more than 50 million agents in Agent 365 since its May 2026 launch. That figure comes from Microsoft itself, with no independent count or breakdown. Still, a figure that size in roughly six months shows why ownership, risk and cost now sit on the IT agenda.
Microsoft now calls Agent 365 an observability platform for AI. Its pitch is that IT, security and business leaders should work from the same inventory of Microsoft and non-Microsoft agents, each team using the console it already knows.
Connected platforms: Entra identities for outside agents
The main change is broader support for what Microsoft calls connected platforms. You connect a supported platform once, and Agent 365 syncs its agents into the registry alongside your Microsoft agents, with no changes to agent code.
The registry records what each agent is, where it was created, who owns it and which tools it uses. On supported platforms, Microsoft says it also shows activity: what the agent was asked, what it did, which models, tools or other agents it called, and whether the task succeeded.
The identity piece matters most. Microsoft's Agent 365 blog says agents from Google, Salesforce, Databricks, Snowflake, Anthropic, OpenAI, and more now get their own identity in Microsoft Entra, so IT and security can govern them alongside Microsoft agents, in addition to Amazon Bedrock agents, which were already supported. With that Entra Agent ID, Microsoft says organizations can:
- Manage each agent's lifecycle
- Apply supported Conditional Access and security policies
- Assess identity risk
- Revoke access when needed
Microsoft names Salesforce Agentforce, Amazon AgentCore, Google Vertex AI, Databricks, Snowflake, Anthropic Claude Managed Agents, OpenAI, Oracle, SAP, ServiceNow and UiPath. Connected platforms are in preview, capabilities vary by platform, and Microsoft says the rollout across platforms will complete in December.
Two setup changes come with it:
- A guided connection wizard walks admins through connecting a platform, tools registry or gateway. That includes the work on both sides: picking what to connect, granting access, and verifying the connection.
- Secretless authentication uses Microsoft Entra workload identity federation, so supported connections don't need long-lived secrets that someone has to create, store and rotate. This applies only to supported connections. It doesn't remove credentials from every integration.
Note the word "supported" in Microsoft's description of policies. An Entra identity for a Salesforce agent does not mean every Entra control behaves the same way inside Salesforce. Consultancy QueryNow makes a sensible suggestion: for an agent on another platform, test what disabling its Entra identity actually stops. Run that test before you count on the revoke option during an incident.
Section summary: Third-party agents can now sit in one registry with Entra identities. The feature is in preview, varies by platform, and rolls out through December.
Shadow AI: finding the agents IT never approved
This part will interest Windows endpoint admins most. Microsoft says Defender and Entra Global Secure Access together can discover more than 60 local AI agents, including Claude Code and GitHub Copilot CLI, along with the MCP servers they're configured to use.
The two products split the work:
- Microsoft Defender finds agents on managed Windows and macOS devices and ties each one to its device and user.
- Entra Global Secure Access adds network-level visibility into each agent's destinations, MCP servers and resources.
- Agent 365 takes those signals and adds local and unapproved agents to the same registry as everything else.
Other reports describe the count slightly differently. WindowsForum's earlier coverage of the Entra announcements reported that discovery expands by 35 local agent types, to 60 in total, including Claude Code, OpenAI Codex, and GitHub Copilot. Microsoft's Agent 365 blog says "more than 60." Treat 60 as a rough floor, not an exact supported list.
Release states differ across this workflow:
| Capability | Status |
|---|---|
| Local agent discovery in Microsoft Defender | Generally available |
| Network-level discovery via Entra Global Secure Access | Preview |
| Shadow AI discovery in the Agent 365 registry | Preview |
| Entra Agent IDs for discovered local agents | Preview |
| Local and shadow agent governance (allow/restrict/block) | Preview |
So the full path from discovery to registry to identity is not generally available. Only the Defender detection layer is.
Separate identities for local agents
A separate preview gives identities to discovered local agents, starting with GitHub Copilot CLI and Claude Code, its own identity, separate from the user's.
This separation matters. Today, a coding agent on a developer's machine mostly acts as that developer, so its file reads, API calls and network traffic look like the person's own activity. Microsoft's answer is agent-specific identity and risk controls that let security teams restrict a risky agent while preserving the employee's own access. As one analysis put it, a separate identity makes it possible to see what an agent did and to switch it off without disabling the employee.
If a prompt-injected coding agent is exfiltrating data, you can contain the agent without cutting off the developer's account.
Enforcement tools on Windows
Microsoft's announcement links discovery to enforcement in Intune, Defender, Purview and Windows:
- Blocking unsanctioned agents: Admins can block unapproved local agents using controls in Intune, Defender and Global Secure Access. This is in preview.
- Device-wide restrictions: Intune can mark devices running unapproved agents as noncompliant, and Entra Conditional Access can then block those devices from company resources. Intune controls for local agents are in preview.
- Containing approved agents: Intune policies set file and network boundaries, and Microsoft Execution Containers (MXC) on Windows enforce them for supported tool execution. Because enforcement happens outside the agent, the agent can't grant itself more access.
- Runtime protection: Defender's runtime protection for local agents is generally available on Windows. It inspects critical points in the agent loop and can block malicious instructions. Microsoft lists Claude Code, GitHub Copilot CLI, OpenClaw and ChatGPT Desktop among supported agents. Admins set Audit or Block mode in the Defender portal or Intune. According to one Ignite analysis, runtime protection for local agents requires Defender for Endpoint P2.
- Data loss prevention: Purview DLP for Microsoft agents and key local agents is in preview. Covered local agents include Claude Code, Claude Desktop, GitHub Copilot CLI, the GitHub Copilot app, and the VS Code extensions for GitHub Copilot and Claude Code.
- Network controls: Global Secure Access adds an MCP Firewall and agent-specific network policies, in preview. Ben Schmark's Ignite recap describes it as enforcing policies on what agents may reach over MCP: servers, tools, resources and methods.
Multi-tenant management, tools registry and activity export
The remaining additions are about running agents across an organization:
- Multi-tenant agent management (preview): Partners and multi-tenant enterprises get one view of Microsoft and non-Microsoft agents across tenants in the Microsoft 365 admin center. Admins can compare an agent's configuration and access in each tenant, review risk, and approve or block agents at scale. For MSPs, an agent decision can be made once instead of once per customer.
- Tools registry (generally available): One inventory of the tools and connections agents use, which helps spot heavily used or risky tools. Connecting external catalogs such as GitHub and Databricks Unity Catalog goes to private preview in December. Azure AI Gateway and LiteLLM gateway support for MCP server governance is a separate preview.
- Export Agent Activity: Agent traces sync continuously into Azure Monitor and Application Insights. Admins filter by agent or platform and choose which prompts and responses to include. Decide that last setting deliberately, because prompts can contain sensitive data.
- Dynamic agent risk (preview): One continuously updated risk level per agent, shown in the Microsoft 365 admin center, Defender, Entra and Purview. Partners including Check Point, Cyera, Darktrace and Zenity feed signals through Microsoft Sentinel.
- Agent Map use cases (preview): Sorts agent sessions by the type of work done, so you can find several teams building the same agent three times.
- Business Analytics (Frontier preview): Usage, spend forecasts and recommendations. Microsoft says analytics for Microsoft services need at least one Microsoft 365 Copilot license in the tenant, and analytics for connected non-Microsoft services need Agent 365 and/or ME7 licensing.
- Action center: Shows inactive and ownerless agents for cleanup. Microsoft's blog says it will reach preview in December. Some partner recaps call it a private preview.
Licensing: what's included and what isn't
Microsoft is widening one entitlement. Qualifying Microsoft 365 subscriptions now get core Agent 365 usage and activity insights for Microsoft pre-built agents and agents built on Microsoft platforms. That covers the agent overview, active users, sessions, exception rates and runtime, and this licensing change is generally available.
The limit is clear, though. Non-Microsoft agent management insights and advanced governance and security still require an Agent 365 or ME7 license. The free tier also gives Microsoft a way to show you what you can't yet govern without paying for the next tier.
What admins should do now
- Map your licensing first. Basic usage data for Microsoft agents comes with qualifying M365 subscriptions. Third-party management and advanced controls need Agent 365 or ME7, and Defender runtime protection reportedly needs Defender for Endpoint P2.
- Turn on Defender local agent discovery. It's generally available and costs little. Even without enforcement, an inventory of Claude Code and Copilot CLI installs on managed endpoints is useful.
- Pilot connected platforms carefully. Start with the platform that touches the most sensitive data. Check what each platform actually supports and test what revoking an Entra identity really stops.
- Choose your prompt-export policy before you turn on Activity Export. Exporting prompts and responses into Azure Monitor creates a new copy of sensitive data in a new place.
- Keep preview features out of production enforcement plans. Network discovery, local agent identities, dynamic risk and multi-tenant management are all in preview.
Bottom line
Microsoft is making a reasonable bet: AI agents are a new kind of insider, and insiders should be governed through identity. Building that on Entra, Defender and Intune, tools admins already run, is a sensible choice. The obvious counterpoint is lock-in. The more third-party agents you register in Agent 365, the more Microsoft's directory becomes the control layer for your Salesforce, Google and OpenAI estate.
What's actually shipping is a solid detection layer on Windows, a generally available tools registry, and a long list of preview features that should look much better by early 2027. For now, inventory everything, pilot carefully, and don't confuse "announced" with "enforced."
References
- Microsoft adds controls for third-party and local AI agents in Agent 365 - Superpower Daily Superpower Daily · 2026-11-20T22:49:38.941000+00:00
- Agent 365 registry for third-party agents: what to do first querynow.com
- What's new in Microsoft Agent 365: Ignite 2026 techcommunity.microsoft.com