I've spent the last few days testing the new Microsoft Entra Agent ID preview in several tenants. The underlying problem has been obvious for a while. Autonomous agents built in Copilot Studio or Azure AI Foundry now work across Microsoft 365 and Azure, and that creates a governance challenge we haven't had before. An agent is no longer just a service principal or an app registration. It can make decisions, access business data and start workflows on its own. Until now, managing identities for these agents meant sorting through a confusing mix of app registrations and service principals. With Entra Agent ID, each agent finally gets its own identity in the directory that you can actually see. That step was overdue.

Agent ID makes agents visible in the Entra admin center

Microsoft announced Entra Agent ID at Build 2025, and it's now in public preview. Agents created in Copilot Studio and Azure AI Foundry receive an Agent ID automatically. You'll find these identities in the Entra admin center under Enterprise applications, which now includes a dedicated agent view listing agents from both platforms. For the first time, we can inventory, monitor and secure agents the same way we handle any other identity.

This matters because agents are quickly turning into first-class actors in the tenant. They don't just help out in the background. They interact with systems, hold permissions, change data and can run critical automations. If these non-human actors don't have clear, individual identities, there's no way to enforce least privilege, ownership or lifecycle management for them. Anyone responsible for compliance or security should see that as a serious gap.

Agent identity closes a gap that workload identities left open

Agent ID follows on naturally from Microsoft's work on workload identities (Workload identities overview). App registrations and managed identities have long been standard for classic app integrations, but agents add more autonomy and more complexity. In the environments I work in, agent governance usually lags well behind governance for users or traditional apps. Who created this agent? Does it have a named owner or sponsor? Which permissions has it collected over time, and does it still need them?

When nobody can answer those questions, agents slip under the radar and never get reviewed, even though they touch sensitive data. Entra Agent ID helps with that. You can:

  • Inventory agents and their owners at a glance
  • Assign and track responsibility (sponsor concept)
  • Enable lifecycle workflows (like removing agents after project end)
  • Apply Conditional Access policies to agents
  • Enforce least-privilege permissions

I've covered the basics and common pitfalls of workload identities in a separate post: Workload identities in Entra ID: Managed identities, app registrations and securing your applications.

Admins can now inventory agents and assign sponsors

The Agent ID preview adds a consolidated agent overview to the Enterprise applications section in Entra. The new agent view lists every agent created through Copilot Studio and Azure AI Foundry, along with technical details, the original creator and, if one has been set, the sponsor. When I looked through several tenants recently, I came across quite a few orphaned or over-privileged agents.

The first thing I'd recommend is taking stock and sorting out ownership, using the new sponsor attribute wherever you can. Microsoft is pushing the sponsor concept in other governance features too, access packages for example. If an agent has no clear owner or sponsor, check whether anyone still needs it, and then either decommission it or hand it to someone new.

Practical tip: Agents without a responsible owner or sponsor are a risk. I schedule regular reviews to find and fix these identities, and where I can, I automate the cleanup with access reviews or lifecycle workflows.

Conditional Access already works for agent identities in the preview

Even at the preview stage, you can apply Conditional Access to agent identities and control their access to sensitive APIs or data. The problems differ from those with human users. MFA and location-based controls generally don't apply, which makes careful scoping of roles and permissions all the more important. I'm hoping to see just-in-time permissions and automated re-certification come to agent identities, as Microsoft outlines in its post on extending Zero Trust to the agentic workforce.

In Copilot Studio, governance controls should be standard practice by now. Restrictive environments, approval workflows, DLP and a clear publishing process aren't optional. I go into practical agent governance in Copilot Studio in more detail here: Copilot Studio governance: risks and controls in 2024.

Each agent gets its identity automatically at creation

The diagram shows how an agent built in Copilot Studio or Azure AI Foundry receives an identity in Entra ID automatically, with both owner and sponsor recorded in the directory:

flowchart TD
    User["User creates agent"] --> CopilotStudio["Copilot Studio/Azure AI Foundry"]
    CopilotStudio --> AgentID["Agent ID in Entra ID"]
    AgentID --> Owner["Owner"]
    AgentID --> Sponsor["Sponsor"]
    AgentID --> EnterpriseApps["Enterprise apps/Agent view"]

Figure: How Copilot Studio and Azure AI Foundry agents receive an identity, including governance attributes.

Agent governance has become a must-have for me

With the Microsoft Entra Agent ID preview, we finally have a practical way to inventory and manage agents in the tenant. That's only the starting point, though. The actual governance work begins now. Owners need to be set, sponsors assigned, permissions limited and lifecycle processes put in place, and IT can't do that without the business teams. The sooner you get agent sprawl under control, the easier it will be to enforce Conditional Access, certificate management and monitoring as the number of agents keeps growing.

If you're already experimenting with agents in Copilot Studio or Azure AI Foundry, don't treat Agent ID as a technical box to tick. You'll need it as the basis for keeping agents from turning into your next governance problem.