1.2.2. Agents as a New Kind of Principal
💡 First Principle: Anything that can act in your tenant needs an identity so you can control and audit it. Agents can act — sometimes autonomously — so Microsoft gives them their own identity type, with sponsors for accountability, and governs them through a registry, much as users are governed through the directory.
Copilot is driven by a person each time. An agent is configured once with instructions, knowledge sources and tools, and then carries out tasks. Some agents act on behalf of a signed-in user, using that user's delegated permissions. Others act with their own access, autonomously, with permissions granted to the agent itself. That second kind is what makes agents a governance problem: a capable actor with standing access that no single person is driving.
Microsoft Entra Agent ID models this with new directory object types. An agent identity blueprint is a template that creates agents but can't access resources itself. Each agent identity is the actual actor that performs the agent's tasks, and it can optionally have an agent user — a user-like account for scenarios that need one. Every agent identity must have at least one sponsor, a person accountable for the agent's purpose and lifecycle, while owners are optional technical administrators.
Microsoft Agent 365 is the governance layer on top. It surfaces Microsoft and third-party agents in an agent registry in the Microsoft 365 admin center, where admins review requests, publish or block agents and manage the tools agents can use, and it connects agents to Entra, Purview and Defender controls for access, data protection and monitoring.
Why the separate identity type? Because an agent's needs don't fit existing objects. A user account assumes a person who signs in interactively. An app registration assumes software that a developer owns. An agent can be created by a business user in Copilot Studio, act on its own, outlive its creator's job role and multiply into hundreds of instances. Without a distinct identity type there is no clean way to apply Conditional Access to agents, make sure every agent has an accountable human, or find all of them when someone leaves.
⚠️ Exam Trap: Don't treat an agent as "just another app" governed only through app consent, or as "just another user" covered by user policies. Conditional Access policies that target users don't apply to agent user accounts — agents have to be targeted as agents.
Reflection Question: A Copilot Studio maker who built a popular agent leaves the company. Using the sponsor concept, explain why the agent doesn't simply become ownerless — and what an administrator should still check.