ooligo

Microsoft Entra Agent ID

identity-and-access-management agent-identity · identity-governance · conditional-access · agent-governance
AI-NATIVE API
RevOpsLegal OpsRecruiting & TACustomer Success
7.0 /10

What it is

Microsoft Entra Agent ID is the part of Entra ID that gives an AI agent its own identity instead of letting it borrow a human’s. Microsoft’s documentation describes it as generally available. It adds four object types to the directory: an agent identity blueprint (the template a family of agents is created from), the blueprint’s principal in each tenant, the agent identity itself, and an optional agent user, a user account for agents that must sign in to systems that only accept users. Every agent identity can carry a sponsor — a named human accountable for its lifecycle and access — separate from its technical owners.

On top of those identities, Entra applies the controls you already run for people: Conditional Access for agents, ID Protection risk detections for agents, access packages and sponsor lifecycle workflows in ID Governance, and a default inactivity policy that finds dormant agent identities, notifies the people responsible, and disables them.

Agent ID is the identity layer, not the inventory. Since 1 May 2026 the agent registry lives in Microsoft Agent 365, and the Entra admin center’s registry blades were retired. Entra issues and secures the identity; Agent 365 lists and governs the agent.

Why ops teams end up here

The question usually arrives as three questions at once: do our ops agents need their own Entra identities, what licence does that take, and what is happening to the Copilot Studio agents Microsoft is migrating?

The first answer is yes, if any agent writes to a system of record unattended. An agent running on its builder’s token shows up in the logs as that person, inherits everything that person can reach, and stops working — or keeps working — when that person leaves. An agent identity gives you a separate audit subject, a sponsor who answers for it, and a policy target.

The third answer is already being decided for you. Message Center notice MC1478488 (24 September 2026) says Copilot Studio agents created before May 2026 may still run on legacy app-registration identities; self-service migration started in Power Platform Advisor on 24 August 2026, and automatic migration of agents without an Agent ID began in September 2026. Migration converts the identity in place, so the agent keeps its client ID and connectors keep resolving. Microsoft Foundry agents get an Agent ID automatically. Agents built outside Microsoft — n8n, Amazon Bedrock, anything in a container — get one through the Entra Auth SDK sidecar or workload identity federation, which exchanges an AWS or Google Cloud token without storing a secret.

Pricing reality

Microsoft’s product page sells Agent ID two ways:

  • Microsoft Agent 365 at $15.00 per user per month, paid yearly.
  • Microsoft 365 E7 at $99.00 per user per month ($90.45 without Teams), which bundles Agent 365 and the Microsoft Entra Suite.

The controls are what you pay for. Microsoft Learn states that Conditional Access for agents and ID Governance for agents each require either E7, or an Agent 365 licence paired with at least Entra ID P1 or Microsoft 365 E3. Notice MC1395007 added two service plans — Entra Conditional Access for Agents and Entra ID Protection for Agents — to Agent 365 and E7 between July and August 2026, and said tenants using these features without one of those licences need to buy one to keep them.

The unit is the human, not the agent. For an E3 shop with 200 employees, Agent 365 adds $15 × 200 = $36,000 a year. For agents acting on behalf of a signed-in user, the user’s own licence also counts: Conditional Access is evaluated against the user’s token and needs E3, ID Protection needs E5, and ID Governance for those agents needs Entra ID Governance or the Entra Suite. A shop that skips the add-on and jumps to E7 pays $99 × 200 = $237,600 a year, which only makes sense if you also want everything else E7 contains.

Best for

An IT or identity lead in a Microsoft 365 tenant at E3 or above whose ops teams already run agents in Copilot Studio, Foundry or n8n. The case where it wins outright: a RevOps or legal-ops agent writes to the CRM or the contract repository with no human in the session, and a compliance review asks which identity made each change, who sponsors it, and which policy blocks it from an unmanaged network.

Not for

A team on Google Workspace, or one whose identity provider is Okta. You would be adopting Entra to govern agents whose users authenticate somewhere else. It is also not the answer for agents that call SaaS APIs with static keys: Microsoft’s own docs say an API key bypasses Entra authentication entirely, so no Conditional Access policy applies.

Versus the alternatives

  • Okta — the other market-share leader in workforce identity. Agent SSO went GA on 24 August 2026 inside core SSO plans at no added cost, issuing short-lived tokens through Cross App Access; Okta for AI Agents is a separate, quote-only subscription. Pick Okta when your people already sign in through Okta and the agents that matter reach non-Microsoft SaaS. Pick Entra Agent ID when the agents are built in Copilot Studio or Foundry, because they are issued an Entra identity whichever IdP you prefer.
  • Rippling AI Governance — the fastest-moving entrant, launched in August 2026 with an MCP gateway and agent identity on top of an HRIS. Pick it when Rippling is the system of record and you want agent access tied to employment status without a separate identity vendor.
  • Zenity, Runlayer, Obot — not identity providers; they inspect what an agent does at the tool call. Add one when the risk is a prompt-injected agent misusing legitimate permissions, which an identity cannot stop.

Watch-outs

  • A permission change can break agents this month. Notice MC1506728 says Entra stopped automatically granting AgentIdentity.CreateAsManager to agent blueprint principals from 4 November 2026, with rollout through late November. Agents that create their own child identities without an explicit admin grant now fail with a 403. Guard: filter Entra audit logs for “Add Service Principal” events with an application actor and an agent identity target, and grant the Graph permission to each blueprint that still needs it.
  • On-behalf-of agents escape agent-targeted policy. A Conditional Access policy that targets an agent identity does not cover requests that agent makes on behalf of a user; those follow the user’s policies, and ID Protection attributes the risk to the user. Guard: write separate policies for the three access patterns — autonomous agents, on-behalf-of flows, and agent users — and test each in report-only mode before enforcing.
  • Two identity types until migration finishes. Copilot Studio agents still on app registrations are not covered by policies written for agent identities. Guard: run the Power Platform Advisor recommendation, migrate a small batch of non-critical agents first, validate channels and connectors, and use the per-agent rollback for anything that breaks.
  • Policy on a blueprint reaches agents created later. That is the point of blueprints, and also a way to block a whole family by accident. Guard: scope new policies to a single agent identity first and widen to the blueprint only after a week of clean sign-in logs.

Related: the enterprise AI rollout stack uses Agent ID as its identity plane, and the agent access governance stack sets which agents may write to the revenue systems, and at what ceiling.