Microsoft Entra Agent ID
Architecture

Plan your agent identity architecture

In brief

The guide now directs cloud and local agent builders to integration tutorials, adds instance mode to architecture planning, clarifies Agent ID provisioning versus ordinary app registrations, and updates client-library and API guidance.

What Entra admins need to know

Administrators should use the updated tutorials and provisioning procedures when designing Agent ID deployments. No required action is stated.

This editorial summary was generated by AI from the documentation changes. Verify important details in the full Microsoft Learn article.

Documentation change

The comparison below shows only the changed extract. Use the full-page view for complete context.

Plan your agent identity architecture

BeforeUse this guide when you integrate're designing identity boundaries for a platform or several separately managed agents. If you want to build a cloud or local agent, start with Choose an integration path. Its tutorials include the required identity setup; this planning guide isn't a prerequisite.

The guide helps you choose the identity type, operation pattern, and blueprint structure that match your AI agent with Microsoft Entra Agent ID, you need to make a series of design decisions. This guide walks you through each decision in order:access and lifecycle requirements.

    Plan the identity boundaries

    For architecture planning, consider:

    • Which identity type your agent needs.
    • Which operation pattern your agent uses, including how it acquires tokens and what context it acts in.
    • How many blueprints your system requires.
    • How many agent identities to create per blueprint.
  • Which instance mode matches that identity relationship.
  • Work through these decisions in order, because earlierEarlier choices shape later ones. Some single-agent deployments might onlyThe selected tutorial sets the instance mode during provisioning; it isn't a separate operating mode your users need the first two steps.to choose. For examples of how these decisions map to real-world agent architectures,architecture examples, see Agent ID design patterns.

    Step 1: Choose an identity type

    • Blueprint-managed credentials and lifecycle: you create, rotate, and delete agent identities through their parent blueprint, not individually.
    • Support for ephemeral agent identities that are created at runtime with inheritable permissions already granted through the blueprint, and deleted when the task completes.

    For a detailed comparison, see Agent identities, service principals, and applicationsAgent identities, service principals, and applications.

    Why not a regular user account?

    • Identity governance processes such as joiner-mover-leaver workflows, access packages, and access reviews aren't designed for agent lifecycle patterns and might incorrectly remove agent access.
    • Your agents would appear in the Global Address List, Teams, and SharePoint alongside human employees, making it harder to distinguish AI agents from people.

    Don't registerUse an appAgent ID for yourthe agent's runtime identity

    If you're used to creating workload identities, your instinct might be to create anAn ordinary app registration or service principal for your agent — for example, by running az ad app create, New-MgApplication, New-AzADApplication, or a POST /applications Microsoft Graph request. Donisn't do this for AI agents. An identity created this way isan Agent ID. Use the supported Agent ID provisioning procedure rather than creating a standard application: it has no sponsor, no agent-specific audit entries, and no blueprint-managed lifecycle. Microsoft Entra doesn't govern it as an agent,application and labeling it an "agent identity"agent. Some supported blueprint requests use /applications with a derived @odata.type; the endpoint alone doesn't make it one.determine the object type. Preserve the request type and headers shown in your tutorial.

    Instead, create anThe agent identity blueprint and then create agent identities from it. Agent identities areis a distinct Microsoft Entra object type (#Microsoft.Graph.AgentIdentity) created throughassociated with a blueprint. Separate management clients, front-end sign-in clients, and downstream APIs can still use ordinary app registrations. Don't replace those objects with the blueprint, not through the application registration APIs.agent's runtime identity.

    Instead of this (standard app registration) Do this (agent identity)
    az ad app create Create an agent identity blueprint, then create agent identities from it
    New-MgApplication or New-AzADApplication Use a supported creation channel with the Agent ID Developer or Agent ID Administrator role
    POST https://graph.microsoft.com/v1.0/applicationsAn ordinary application creation request Grant AgentIdentityBlueprint.CreateFollow the typed blueprint request in your scenario tutorial, then create the blueprint and agent identitieswith that procedure's management permissions.

    In .NET,For a cloud ASP.NET Core agent, use the standard Microsoft.Identity.Web client APIs in Call a custom API as an agent. For a hosting platform with a fleet, use Microsoft.Identity.Web.AgentIdentities package: call builder.Services.AddAgentIdentities(), then acquire tokens with WithAgentIdentity(...)and explicit child-identity selection in the fleet tutorials. See Call custom APIs from an agent.For a local public client, use MSAL without a client credential.

    For a full comparison, see Agent identities, service principals, and applications.

    | Consent model | Admin consent required | User or admin consent | | Token subject | Agent identity | User (with agent as actor) | | Common scenarios | Background processing, scheduled tasks, system-to-system | Chat assistants, user-facing copilots, agents acting on user data | | Access scope | Broad, tenant-wide accessPermissions granted to the agent, further constrained by the resource's authorization model. | Scoped toDelegated permissions and the signed-in user's dataaccess, further constrained by the resource's authorization model. |

    Choose autonomous when your agent:

    • Runs background or scheduled tasks.

    Some agents need both. For example, an agent might run a nightly background sync using the autonomous pattern and also respond to user chat messages using the interactive pattern. In this case, implement both OAuth flows and select the appropriate token based on the operation.

    Step 3: Decide how many agent identity blueprints

    An agent identity blueprint managesdefines authentication settings for all theits agent identities created from it. Theand holds credentials for confidential clients. A multi-instance blueprint holds theuses those credentials used to acquire tokens on behalfimpersonate its child identities. A compromise of those agent identities. Because of this capability, a blueprint compromiseshared credentials can affect every agent identity under it.that uses them. Choose your agent identity blueprint count based on your security boundaries.

    Default: use one blueprint per trust boundary.

    • Scale-out or replicas: Running multiple instances of the same agent doesn't require multiple blueprints.
    • Memory or context separation: Agent memory is typically a shared data store filtered by session ID at retrieval time, which doesn't require separate blueprints.

    For more information about blueprints, see Agent identity blueprintsAgent identity blueprints.

    Step 4: Decide how many agent identities per blueprint

    Multi-agent architectures where multiple agent identities are appropriate can include sequential pipelines with distinct agent roles and concurrent orchestrations with specialized domain workers. For more information, see AI agent orchestration patterns.

    Step 5: Choose the instance mode

    Choose the blueprint's mode before creating it. Single-instance mode represents one Agent ID per tenant, with the same OAuth client ID as the blueprint and no blueprint principal. Multi-instance mode supports separate child Agent IDs with distinct client IDs.

    Registered public-client agents must use single-instance mode. Confidential agents can also use it when one identity is appropriate. If the blueprint needs to provision identities with independent permissions or lifecycles, use multi-instance mode.

    Having only one child identity doesn't make an existing multi-instance blueprint single-instance. The setting is fixed at creation. For details, see Blueprint instance modes and Create a single-instance agent identity.

    Related content

    Daily Entra.News

    Get daily email updates

    Get a concise summary of the latest Microsoft Entra updates delivered straight to your inbox.

    Loading the secure signup form…