top of page

The Entra Agent Toggle Closes the Future, Not the Past

Kostas Tsiolas
Oct 28
5 min read

From 4 November 2026, any user without an Entra role can create agent identity blueprints. Switching that off takes two minutes. Governing what already exists is the real work.

On 13 October, Microsoft published two Message Center notices about AI agent identities in Microsoft Entra. Both start rolling out on 4 November 2026.

Read together, they pull in opposite directions.

One loosens who can create the templates that AI agents are built from. The other tightens what those templates can do on their own.

Most commentary so far has focused on a single action: switch the new setting off. That advice is correct. It is also incomplete. What changes on 4 November

Notice

What changes

After rollout, unless you act

MC1506733

Default users (no built-in or custom Entra role) can create agent identity blueprints

Allowed

MC1506728

Entra stops automatically granting AgentIdentity.CreateAsManager to agent blueprint principals

Explicit admin consent required

Rollout begins on 4 November and is expected to complete by late November 2026.

An agent identity blueprint is the template an AI agent's identity is created from. It carries the configuration that every agent identity built from it relies on. Microsoft requires each blueprint to have a sponsor, the person or group accountable for the agent, and recommends an owner, who can change it.

Keep those two words in mind. They are where this gets interesting.

Step one: the toggle

Sign in to the Microsoft Entra admin center and go to:

Entra ID > Users > User settings > Default user role permissions

Set Users can register agent identity blueprints to No, then save.

A naming note: Microsoft's Message Center calls this setting "Users can create agent blueprints". The admin center label reads "Users can register agent identity blueprints", directly below "Users can register applications". Search for the portal label.

What you need to know:

  • You can set it now. Your choice applies when the rollout reaches your tenant.

  • Approved builders keep working. With the toggle on No, supported administrators and users holding the Agent ID Developer or Agent ID Administrator role can still create blueprints.

  • It is independent of app registrations. Review "Users can register applications" separately.

In our own tenant, it is already set to No. To verify at scale, the setting is exposed in Microsoft Graph as defaultUserRolePermissions.allowedToCreateAgentIdentityBlueprints. At the time of writing it appears only on the beta endpoint.

Connect-MgGraph -Scopes Policy.Read.All
$policy = Invoke-MgGraphRequest -Method GET `
  -Uri "https://graph.microsoft.com/beta/policies/authorizationPolicy/authorizationPolicy"
$policy.defaultUserRolePermissions.allowedToCreateAgentIdentityBlueprints
# False = default users cannot create agent identity blueprints

What the toggle does not do

This is the part most guidance skips.

According to MC1506733, the setting does not:

  • change or remove existing agent blueprints

  • change who can create agent identities from existing blueprints

  • block an app or an agent from creating identities

  • change creation requirements in platforms such as Microsoft Foundry or Security Copilot

And per Microsoft's documentation, the owner of a blueprint can create agent identities from it without any Agent ID role.

Put those together.

Any blueprint a user created during a pilot, a proof of concept, or "just trying it out" keeps working after you flip the toggle. Its owner keeps the ability to create new agent identities from it.

The toggle will not surface any of them. Only an inventory will.


Step two: the permission change most teams will miss

MC1506728 lands on the same day and gets far less attention.

Today, every agent blueprint principal receives the Microsoft Graph application permission AgentIdentity.CreateAsManager automatically, without administrator approval. That permission lets a blueprint create agent identities with no signed-in user.

From 4 November, that automatic grant stops. An administrator must grant it explicitly where an agent genuinely needs it.

What this means in practice:

  • Who is affected: agents that create identities on their own (app-only) and have no other permission allowing it.

  • How to check: Entra ID > Agent blueprints > select the blueprint > Granted permissions. If nothing else authorises identity creation, the agent depends on the automatic grant.

  • What breaks: after enforcement, app-only creation requests that rely only on the removed grant fail with a 403 Forbidden error.

  • What to do: explicit grants have been possible since 16 October, through tenant-wide admin consent or a manual app-role assignment. Grant only where needed, and record why.

This one is good news for governance. A permission that was granted silently to every blueprint becomes visible, deliberate and revocable.


Why this is a governance question, not a settings question

In almost every tenant I review, nobody checks third-party app permissions, leftover consents from retired tools, or service-principal sign-in logs. Security effort goes into human identities. Non-human identities accumulate quietly, and nobody owns them.

App registrations followed exactly this path. Agent identities are next, with one difference: agents are built to act.

That's why sponsor and owner matter more than the toggle.

The sponsor answers "who is accountable for this agent?" The owner answers "who can change it, and create more identities from it?"

In a governance review, those are the first two names a CISO should ask for. In most environments, they are also the first two nobody can produce.

Pre-4 November checklist

  1. Set the toggle. "Users can register agent identity blueprints" to No. Verify via Graph if you manage several tenants.

  2. Inventory existing blueprints. Entra ID > Agents > Agent blueprints. Export every blueprint with its owners and sponsor.

  3. Challenge every owner without a role. Any blueprint owned by a user with no Entra role deserves a conversation. Confirm it is still needed, then confirm the owner should remain an owner.

  4. Validate sponsors. Is the sponsor still in the organisation, in the right role, and aware they are accountable?

  5. Decide who builds agents. Assign Agent ID Developer to the people who actually build agents. Review who holds Agent ID Administrator as carefully as Application Administrator.

  6. Check CreateAsManager dependencies. For each blueprint, review Granted permissions. Grant AgentIdentity.CreateAsManager explicitly only where app-only identity creation is required, and document the reason.

  7. Close the app-registration gap too. If "Users can register applications" is still Yes, the agent conversation is a good moment to fix it, and to review existing app permissions and consents.

  8. Watch for new activity. In the audit logs, filter Category: ApplicationManagement and Activity: Add service principal. Look for entries initiated by an application where the target is an agent identity.


A note on scope

Microsoft's notices describe a worldwide rollout and do not mention GCC, GCC High, DoD or other sovereign clouds. If your tenant runs in one of those environments, check your own Message Center rather than assuming the same dates.


Three questions for leadership

  • Who in our organisation is allowed to create an AI agent identity, and who decided that?

  • For every agent that exists today, can we name its sponsor and its owner?

  • If an agent misbehaves tomorrow, how quickly can we identify which identity it used and who approved it?

If those answers take more than a day to assemble, the toggle is not your biggest gap.


The bottom line

AI agents will arrive in your tenant whether or not you plan for them.

The toggle decides how they arrive from now on. The inventory decides whether you know what is already there.

If you would like a second pair of eyes on your agent and application identity inventory, Nimbus Cyber runs identity governance reviews for Microsoft Entra environments. It starts with a conversation: nimbuscyber.eu/services.

Comments


bottom of page