Every admin team running Copilot Studio eventually asks the same question: where do we actually govern these agents? On November 11, 2026, Microsoft’s Copilot Studio blog answered it directly in Where do I govern my Copilot Studio agents? The short answer is five places, organized by outcome rather than by admin center, with a lead role for each.
The timing matters. Supported Copilot Studio agents can now have their own Microsoft 365 account, a directory profile and a place in the org chart, and can join Teams and Outlook with scoped permissions. Microsoft’s framing is blunt: makers see a digital teammate, while admins see a new actor whose access, data use, cost and behavior all need governing. The framework is a good map. Before you build a governance model on it, it is worth checking each of the five surfaces against what Microsoft Learn currently documents, because the coverage is not as even as the diagram suggests.
The five outcomes at a glance
| Outcome | Lead role | Where | Main controls named in the blog |
|---|---|---|---|
| Enable makers safely and at scale | Power Platform admin | Power Platform admin center | Environment routing, environment groups and rules, advanced connector policies, sharing controls |
| Configure agent identity and resource access | Identity and access admin | Microsoft Entra | Entra Agent ID, Conditional Access for agents |
| Protect the information agents use and produce | Compliance admin and data security team | Microsoft Purview | Data Security Posture Management for AI, classification, sensitivity labels, data loss prevention, audit |
| Maintain estate-wide observability and prove business value | AI admin | Microsoft Agent 365 and the Microsoft 365 admin center | Agent registry, agent map, agent details, tool governance, Copilot Credit cost management |
| Detect and respond to threats | Security admin and SOC analysts | Microsoft Defender | Agent security posture, threat investigation, real-time protection |
Two framing points from the blog are worth carrying into any governance model. First, Agent 365 complements the Power Platform admin center rather than replacing it, and Copilot Studio data policies complement Purview information protection rather than replacing it. Second, Microsoft recommends starting from the outcome you need, not the admin center you expect to use.

1. Enable makers: sharing controls stop new sharing, not existing access
The Power Platform admin center still sets the ground rules: environment routing into personal developer environments, environment groups and rules, an allowlist approach through advanced connector policies, and sharing controls.
The sharing controls page (last updated November 5, 2026) is precise about their limits. Rules are enforced when someone tries to share and do not affect existing access. They can take up to an hour to take effect. They apply only to agents that require authentication. The blog says sharing controls help limit agent user account risk, but the Learn page does not mention agent user accounts at all. If you are relying on sharing limits to contain agent accounts, treat that as unconfirmed and review agents already shared before you tighten the rules. The Managed Environments setting that decides your blast radius is still the starting point.
2. Identity: one agent can need three Conditional Access policies
The blog distinguishes three identity patterns: an agent identity, an agent acting on behalf of a user, and an agent user account. The Conditional Access for agents page (last updated September 30, 2026) shows why that distinction matters. Each pattern has a different token subject and a different policy target.
| Access pattern | Token subject | Conditional Access targets |
|---|---|---|
| On behalf of a user | The signed-in user | Users and groups |
| Application-only | The agent identity | The agent identity or its blueprint |
| Agent user account | The agent’s user account | The “Agents” option, not the Users assignment |
A policy that targets the agent identity does not cover the agent’s user account, and the reverse is also true. You cannot scope agent user accounts by group membership, and Conditional Access does not apply when an agent authenticates with an API key. Licensing is also specific: Microsoft 365 E7, or Microsoft Agent 365 plus at least Entra P1 or Microsoft 365 E3.
The Entra Agent ID management page mentions Copilot Studio only once, as a product-level control for blocking new agent identities. It does not say whether Copilot Studio agents appear in the agent identity lists or fall under ID Protection. If you have not moved your agents yet, Copilot Studio Entra Agent ID Migration: Do It Before Microsoft Does It For You covers that step.
3. Information protection: check which agents Purview actually covers
The blog itself says Purview coverage varies by agent and interaction type, and the documentation confirms it. The Purview page for AI agents covers Agent 365 agent instances and treats them like users in most policies. It does not address Copilot Studio agents or agent user accounts directly.
For the agents it does cover, three details change how you design policy. Audit and classification are automatic, but data loss prevention and Insider Risk Management need the agent added explicitly, the same way you would add a user. A DLP block can break an agent’s workflow because the agent does not know it was blocked, so owners have to watch those policies. And sensitivity labels do not carry over to content the agent creates, so new output is not labeled or encrypted automatically. Copilot Studio data policies sit alongside all of this as a platform control; Power Platform Data Policies vs. Microsoft Purview DLP explains where each one stops.
4. Observability: the registry may show one agent more than once
Agent 365 and the Microsoft 365 admin center are where the AI admin sees the whole estate: the agent registry, the agent map, the agent details page, tool governance, and Copilot Credit cost management with spending policies and alerts.
The registry page (last updated October 23, 2026) has details that will surprise anyone using it as an inventory. A Copilot Studio agent can appear more than once, under different publisher types, by design. Publishing in Copilot Studio alone may not make an agent visible in the admin center. Removing one entry does not remove the others. Risk signals in the registry need an E7 or Agent 365 license, and risk counts can lag the security portals by up to an hour. The registry does include an “Agents without owners” filter, which is the most useful starting point for a first review. For why Agent 365 now sits at the center, see Copilot Studio’s security page now names Agent 365 as the control plane, and for spend controls, Zero Credits Allocated Does Not Mean Zero Spend.
5. Threat protection: Copilot Studio support is in preview, and the default only audits
The blog lists jailbreaks, prompt injection, misuse of overly broad tools and unusual activity spikes as the threats Defender addresses. The real-time protection page (last updated October 15, 2026) is narrower and more specific.
- Copilot Studio agent support is labeled preview, and Copilot Studio must be connected to Defender first.
- The built-in default rule audits all agents and records behaviors without blocking. Blocking needs a custom rule.
- Only agents with a Microsoft Entra agent ID appear when you scope a rule, which ties threat protection back to the identity work in outcome 2.
- The documented scenarios are malicious content propagation, secret exfiltration, evasion techniques and email to untrusted domains. Prompt injection and jailbreak are not named as scenarios on that page.
- A blocking rule suppresses near-real-time alerts for the agents it covers, and prompt snippets are collected as alert evidence by default.
What the framework gets right
The most useful part of the blog is not the product list. It is the operating advice at the end. Assign a lead for each outcome. Agree in advance on the decisions that need more than one admin: moving an agent to production, granting it new resource access, and letting it reach a wider audience. Then review ownership, access, data protection, usage and risk on a regular schedule. That turns five admin centers into one governance process instead of five separate ones.
The gaps above do not undermine the framework. They show where it is still ahead of the documentation, and where licensing quietly decides how much of it you can use. Several of the controls that matter most for agent accounts, including Conditional Access for agents and registry risk signals, need Agent 365 or Microsoft 365 E7.
A checklist before you adopt the framework
- Name a lead for each of the five outcomes, and write down who signs off when an agent moves to production, gains new resource access, or reaches a wider audience.
- Confirm which of your licenses cover Conditional Access for agents and registry risk signals before you design policies that depend on them.
- Build Conditional Access policies per access pattern: on behalf of users, agent identities and agent user accounts each need their own.
- Review agents that are already shared before tightening sharing controls, since new rules do not remove existing access.
- Add agents explicitly to Purview DLP and Insider Risk policies where they are supported, and assign someone to watch for silent DLP blocks.
- Start your registry review with the “Agents without owners” filter, and expect duplicate rows for Copilot Studio agents.
- Connect Copilot Studio to Defender, run the default audit rule first, and add blocking rules only for high-confidence scenarios.
Microsoft has drawn a clear map of where agent governance lives. The work for admin teams is to put a name on every outcome, and to check each control against the documentation before trusting it to cover agent accounts.
Sources
- Where do I govern my Copilot Studio agents?, Microsoft Copilot Studio Blog, November 11, 2026
- Copilot Studio sharing controls, Microsoft Learn
- Conditional Access for agents, Microsoft Learn
- Manage agent identities, Microsoft Learn
- Purview protections for Agent 365 AI agents, Microsoft Learn
- Agent registry in the Microsoft 365 admin center, Microsoft Learn
- Real-time protection for AI agents, Microsoft Learn
