BlogAI
How Do I Get My AI Agents into Agent 365?
Video · 5:16
The 5-minute version
If you had users in your environment and didn't know who they were, what they were doing, or what they were accessing, you'd treat that as an emergency. I think the same is true for agents. A common question I get about Agent 365 is how to connect the various types of agents for security and governance, and the answer depends on the type of agent. The businesses I talk with are asking about the agents they have today and also the ones they're building or plan to add.
Agent 365 gives you one registry of the agents in your environment and connects each one to the tools you already use to secure your users. Microsoft Entra handles the agent's identity, Microsoft Purview handles what happens to data, and Microsoft Defender watches for threats. The idea is to treat an agent like a user, so you know who it is, what it can reach, and what it's doing.
Agent 365 is licensed per user, on its own or as part of Microsoft 365 E7, as of November 2026.
What are the four types of agents?
Generally, you can group agents into four types: first-party agents from Microsoft, third-party agents from a vendor, agents on a platform that registry sync supports, and custom agents your developers write. How an agent gets connected depends on who built it, where it runs, and what data it's trying to access. In this order, the four types go from connected out of the gate to work your developers have to do.
| Agent type | How it's registered with Agent 365 | Who enforces access to data |
|---|---|---|
| First-party: agents from Microsoft, including Microsoft 365 Copilot, Copilot Studio, and Microsoft Foundry | Automatically | Microsoft Entra and Microsoft Purview, using the user's permissions, the agent's Entra Agent ID, and the policies you already manage |
| Third-party: a vendor's agent, such as the ones from Salesforce, HubSpot, or Workday | Like an app today, with sign-in, install controls, and consent. Vendors are moving toward Entra Agent ID. | The vendor's platform for its data, and Microsoft 365 for yours |
| Registry sync: agents on another supported platform | You connect the platform and its agents sync in, read-only today | The platform where the agent runs |
| Custom: agents your developers write | Your developers integrate the Agent 365 SDK and CLI into their software development lifecycle | Entra Agent ID and the permissions your security team approves |
How do first-party agents connect?
First-party agents connect on their own, because they're built on Microsoft's platforms. That covers Microsoft 365 Copilot and its built-in agents such as Researcher and Analyst, agents your people build in Microsoft Copilot Studio, and agents your developers build in Microsoft Foundry.
Microsoft's own agents appear in the Agent 365 registry automatically. Agents built in Copilot Studio and agents published from Foundry are registered automatically too, and each new one gets its own Entra agent identity. An agent identity is the agent's own identity in Entra, with its own permissions and its own audit trail.
Once an agent is registered, the tools you already run for your users cover it. Microsoft Purview covers Microsoft 365 Copilot with everything it offers for AI, including sensitivity labels, data loss prevention, auditing, and eDiscovery. Most of that extends to Copilot Studio and Foundry agents, and the details depend on how each agent is built. Microsoft Defender lists Copilot Studio and Foundry agents in its AI agent inventory and watches them for threats, and that coverage is part of the Agent 365 license.
How permissions get assigned depends on the platform. Microsoft 365 Copilot works with the access of the person using it, so it only returns content that person can already open. A Copilot Studio agent's permissions come from the connectors it's allowed to call, and most connectors run with the user's own permissions. When a maker submits a Copilot Studio agent for the whole organization, an admin reviews the request, picks who gets it, and approves its permissions. A Foundry agent's identity starts with no permissions, so a developer or an administrator grants what it needs.
Microsoft automates most of this because it controls these platforms end to end. Your first-party agents are registered into the Agent 365 inventory without anyone having to connect them manually, so your team's focus is on ensuring least privilege access for each agent.
How do third-party agents connect?
Most third-party agents connect the way the vendor's app does today, while the trend is for vendors to adopt agent identity in their agents. Microsoft already lists partner agents that are integrated with Agent 365. Capabilities will increase over time as more vendors take the agent identity approach, which gives security teams more granular control over what those agents can access in the environment. It's an evolution, and the app controls you have today still apply while vendors make that move.
A vendor's agent typically exists so people can get to the information in that vendor's platform from where they already work. Someone asks a question in Microsoft 365 Copilot or Microsoft Teams, and the answer comes from the data in the vendor's platform, such as Salesforce, HubSpot, or Workday. Some vendor agents also reach back into Microsoft 365 to read email, Teams messages, or files for context. When a vendor publishes its agent into Teams or Copilot, it shows up in the Agent 365 registry as an external partner-built agent, and an admin controls who can install and use it.
A common question I get is whether you enforce access to the data in Entra or in the vendor's platform. The platform that holds the data is responsible for enforcing what comes back, based on the user's and the agent's permissions in that platform. So the vendor's platform decides what its agent returns from its own data, and Microsoft 365 enforces access for the data stored in Microsoft 365. When a vendor's agent reads Microsoft 365 data on a user's behalf, it can reach only what that user can already reach. It also depends on how the agent signs in to the vendor's platform: when each person's identity is passed through, the vendor can apply that person's permissions, and when the agent connects with one shared service account, the vendor sees the same identity for everyone.
I ask three questions about any third-party agent:
- Does it access Microsoft 365 data, or other data outside its own platform?
- Where is it used: in the vendor's app, or as the vendor's agent in Teams or Copilot?
- How do people sign in: with Entra single sign-on, or separately?
The answers tell you which controls are yours and which are the vendor's. Many third-party agents only work with data in their own platform, so governance and enforcement stay in that platform, and Microsoft's part is sign-in and who can use the agent. What you get is a clear view of which vendor agents are in your environment, who can use them, and what Microsoft 365 data they've been approved to reach.
What is registry sync?
Registry sync is a capability in Agent 365 that brings agents running on another supported platform into your Agent 365 registry, so you see them beside everything else. These are the platforms Microsoft's documentation lists as supported.
| Supported platform | Vendor |
|---|---|
| Amazon Bedrock | Amazon Web Services |
| Google Vertex AI | Google Cloud |
| Agentforce | Salesforce |
| Databricks Genie | Databricks |
| Claude Managed Agents | Anthropic |
| Oracle Generative AI Agents | Oracle |
| Snowflake Cortex | Snowflake |
This list is as of November 12, 2026, and Microsoft has been adding platforms.
You connect the platform to Agent 365, and the agents running there show up in your registry. The connection is read-only today. Agent 365 doesn't run or change those agents, and the sync doesn't give them an Entra identity, so access to data is still set and enforced on the platform where the agent runs. The initial capability of registry sync is observability, which means knowing what agents you have and where they run.
The benefit is one inventory of your agents, including the ones that live outside your tenant.
How do custom agents connect?
A custom agent is one your team writes itself, on whatever model, framework, and host fits the job, so nothing registers it for you. Your developers make the connection part of how they build.
Microsoft's Agent 365 SDK allows your developers to give the agent an Entra-backed identity, and Microsoft's observability library reports what the agent does. The Agent 365 CLI creates the blueprint (the template the agent's identity is made from), creates the agent identity, and registers the agent in Agent 365. A Global Administrator then approves the permissions it asked for. That keeps development and security separate, the way we've handled applications for years: the developer builds the agent, and the security team decides what it can reach.
If you want data governance inside the agent as well, Microsoft Purview publishes a set of APIs so your developers can build enforcement into the agent's workflow. The agent sends a prompt or a response to Purview, Purview returns its decision based on your policies, and the agent acts on that decision, for example by blocking a prompt that contains sensitive data. Defender sees what the agent reports through Agent 365.
I'd build those steps into your software development lifecycle, so an agent goes through your security review before its identity is given access. Custom agents take the most work of the four to deploy and manage, but they're custom for a reason, and it's the traditional buy-or-build conversation we've had for decades. I covered the details in How Do You Secure a Custom AI Agent's Access with Agent 365? and Agent 365: How Agent Blueprints Work.
My take and next steps
I think Agent 365 is largely a conversation about what agents you have and how you connect each one into the tools you already use for your users. It sounds complex, but it's the same way we manage users.
If I were starting, I'd list the agents in the business and sort them into these four types, because the type tells you how much work it takes to connect and who enforces access to the data. Then I'd pick a handful that represent the rest and test three things on each: can you secure its access to your environment, can you observe that access, and can you enforce your data policies on what it does.
Where this doesn't apply is the AI tools people install on their own devices and the AI sites they visit. That's handled more on the device, in the browser, and on the network, with tools like Microsoft Defender for Endpoint, Microsoft Entra Internet Access, and Microsoft Purview Data Security Posture Management for AI, and I'll be covering those specific angles in future posts.
I help organizations with Agent 365 proof of concepts, pilots, and production rollouts.
Working through the same problems?
Jason compares notes with IT leaders every week. Reach out on LinkedIn.