3.2.4. The Agent Registry: Discover, Review, and Publish
💡 First Principle: You can't govern agents you can't see. The registry's job is to be the single inventory of every agent — Microsoft's, partners', your developers' and your users' — and then to be the gate through which an agent becomes available to others.
Discover. Agents > All agents > Registry lists agents from Copilot Studio, SharePoint, Agent Builder, Microsoft Foundry, the Microsoft 365 Agents Toolkit, Microsoft's own agents (such as Researcher) and non-Microsoft platforms. Publisher types are Microsoft, external partner-built, published by your org (admin-approved line-of-business agents) and shared by creator. Summary tiles show total agents, agents without owners, unmanaged agents (created or managed outside Agent 365, without its protection and observability) and agents at risk. Filters cover status, publisher type, channel, platform and data source. Viewing the registry requires a role, not a license.
Two details are easy to miss:
- The same Copilot Studio agent can appear twice — once as shared by creator and once as published by your org. That's by design, and removing one entry doesn't remove the other.
- Registry sync (Connected platforms) brings in agents from platforms such as Amazon Bedrock, Google Vertex AI and Salesforce Agentforce. Connections are read-only, synced manually (there's no schedule), and use key-based credentials.
Review and publish. Agents > All agents > Requests holds agents submitted for approval (from Copilot Studio, Foundry and the Agents Toolkit) in states such as Pending review, Pending update and Pending activate:
- Publish to store — choose who can install it, optionally who gets it preinstalled, apply a policy template, review and consent to its permissions, then publish.
- Reject submission — from the request's menu, or in bulk with the rule that rejects requests older than N days.
- Pending update — until you approve the update, users keep the previous version.
- Pending activate — approval required before users can create agent instances; the approver scopes the audience, applies a template and grants consent.
Only AI Administrators and Global Administrators can approve requests.
⚠️ Exam Trap: User requests for a blocked agent work only for Microsoft-built agents — they appear as Allow user to install, and the admin must unblock the agent before approving. Users can't request third-party agents through the admin center, and not even a Global Administrator can approve such a request there.
Reflection Question: A developer pushes version 2 of an approved agent with new Graph permissions. What do users see until you act, and what must you review before approving the update?