85 agentic tools audited for security incidents · Talk to the lab →
MotionticBook a call
Glossary

What is agent identity?

Agent identity is a distinct, verifiable identity issued to an AI agent, used to authenticate it, authorise what it can access and attribute every action it takes, instead of a shared service account or a borrowed human login.

Checked

Early agents ran on whatever credentials were at hand: a developer's API key, a shared service account or the user's own session. That makes two questions hard to answer after an incident: which agent did this, and why was it able to? Agent identity treats each agent as its own principal, with its own credential, permissions and audit entries.

In 2026 the major identity platforms shipped this as a product. Google describes Agent Identity as the foundation layer of its agent governance stack, assigning every agent a unique cryptographic ID, and Microsoft Entra Agent ID models each agent as a special kind of service principal.

How Google implements it

  • A SPIFFE ID per agent

    Each agent gets an identity based on the SPIFFE standard, tied to its lifecycle and mapped to the resource where it runs. Permissions are granted to that identity directly in IAM.

  • Short-lived certificates

    Google Cloud provisions an X.509 certificate for each agent, valid for 24 hours and kept current automatically.

  • Bound tokens

    Access tokens are cryptographically bound to the certificate, using mTLS with Google APIs and DPoP across Agent Gateway, so a stolen token cannot be replayed outside the agent's runtime.

  • Not a service account

    Agent identities are not shared between workloads by default, cannot be impersonated and cannot have long-lived keys. When an agent acts for a user, audit logs show both.

  • Auth manager

    A credential vault and broker for outbound tool access, handling API keys and OAuth, including delegated end-user tokens.

How Microsoft implements it

  • Blueprints

    Every agent identity in Microsoft Entra is created from an agent identity blueprint that defines the kind of agent, so an administrator can apply a Conditional Access policy to, disable, or revoke a permission for every agent of that kind at once.

  • No credentials of its own

    The blueprint holds the credentials and acquires tokens on the agent identity's behalf.

  • A sponsor

    Each agent identity can record the person or group accountable for it, for example to contact during a security incident.

  • An optional user account

    For systems that only accept user accounts, an agent can be given a linked agent's user account, always one to one with its agent identity.

Acting for itself or for a user

An agent can act on its own authority, as a background job does, or on behalf of a signed-in person, and the two need different tokens. In Microsoft's model, an autonomous agent acquires app tokens for itself, while an agent called with a user's token acquires user tokens in which the user is the subject and the agent is recorded as the actor. Either way, the log answers who asked and which agent acted.

OWASP's guidance on excessive agency points the same way: run tools in the user's context with minimal scope, rather than through one generic high-privilege identity.

Lifecycle gaps to watch

  • Stale grants

    Google notes that deleting an agent does not remove IAM bindings that reference it; they must be cleaned up when the agent is retired.

  • Redeploys get new identities

    A replacement Google agent gets a new resource ID and so a new principal, even with the same name and code. Permissions defined in infrastructure code survive this; permissions granted by hand do not.

  • Orphaned agents

    Agents that outlive their owner's attention are a governance gap; Microsoft's sponsor field exists to close it.

How Motiontic's engineers apply it

Motiontic's engineers give each production agent its own identity on the client's platform rather than reusing a team service account or an engineer's key, with permissions declared in infrastructure code so they survive redeploys. Where an agent acts for a user, the user's delegated token is used and both names reach the audit log, which is what compliance reviews ask for.

Agent identity compared with older approaches

Agent identity compared with older approaches
Shared service accountBorrowed user loginAgent identity
What the log showsThe service, not which agentThe user, not the agentThe specific agent, and the user when acting for one
PermissionsUnion of every workload's needsEverything the user can doScoped to the one agent
CredentialOften long-lived keysUser session or passwordShort-lived and platform-issued
Revoking one agentBreaks every workload sharing itMeans revoking the userAffects only that agent

Common questions

What is an AI agent identity?

It is an identity issued to one AI agent, with its own credential and permissions, so systems can authenticate the agent, limit what it can reach and record what it did.

Why not run an agent on a service account?

Service accounts are often shared and carry long-lived keys, so a log entry cannot say which agent acted and revoking one agent breaks others. Google's agent identities cannot be shared by default, impersonated or given long-lived keys.

What is Microsoft Entra Agent ID?

It is Microsoft's identity layer for AI agents. Each agent identity is a special service principal created from a blueprint, with no credentials of its own and an optional accountable sponsor.

What is a SPIFFE ID?

A SPIFFE ID is a workload identity string defined by the open SPIFFE standard. Google Cloud assigns one to each agent, in the form spiffe:// followed by a trust domain and the agent's resource path.

Running agents that have to hold up?

Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.