IT Glossary

AI Agent Identity

AI agent identity gives each autonomous AI agent its own owner, credentials and scopes. How it works and how it differs from a service account.

November 2, 2026

What is AI agent identity?

AI agent identity is a type of non-human identity that represents one autonomous AI agent: the named human who owns it, the credentials it authenticates with, the scopes it holds and the systems it can act in. It lets IT and security answer, for every action an agent takes, which agent acted, on whose behalf and with which rights.

Key takeaways

  • An AI agent identity ties one autonomous agent to a named human owner, its credentials, its permission scopes and the systems it may act in.
  • An AI agent identity differs from a service account because the agent chooses its own actions, chains tools and often uses a user's delegated permissions.
  • Every AI agent identity should combine least-privilege scopes and short-lived credentials, with activity logged and the identity removed when the agent is retired.
  • The MCP specification defines optional OAuth 2.1-based authorization for HTTP transports, so an AI agent identity often holds OAuth tokens rather than passwords.

How AI agent identity works

  • Registration: each agent gets its own identity, for example an OAuth client or a workload identity, never a shared or personal human account.
  • Ownership: a named person is accountable for the agent's purpose, its access and its retirement.
  • Credentials: the agent authenticates with OAuth client credentials, delegated user tokens, API keys or a workload identity; short-lived tokens beat static keys stored in code.
  • Scopes and delegation: permissions match the task, such as read-only access to a ticketing tool. When the agent acts for a user, its token carries that user's rights within the granted scopes.
  • Logging and offboarding: actions are recorded under the agent's identity, and its credentials are revoked when the agent is retired or its owner leaves.

How is an AI agent identity different from a service account?

  • Service account: runs a fixed script or integration. Its actions are predictable, so a static permission set and a periodic review are usually enough.
  • AI agent identity: backs software that decides what to do next, calls several tools in a row and may act with a user's permissions. The same rights can produce very different actions, so scopes must be narrower and logs more detailed.

Two neighboring terms are often mixed up. Non-human identity is the category that includes service accounts, API keys, workloads and agents. AI agent governance is the program around agents: inventory, risk assessment, approval and review. Protocols are starting to formalize agent access: the MCP authorization specification defines an optional authorization mechanism for HTTP-based transports, based on OAuth 2.1 (an IETF draft) and related OAuth RFCs.

Examples and use cases

A 150-person marketing agency runs three AI agents: one triages support tickets, one drafts client reports from the CRM and one updates the project tool. All three were connected with the personal API keys of the employee who set them up. When she leaves, offboarding disables her account in the identity provider, but the keys keep working, and nobody can tell which actions the agents took for whom. The fix is one identity per agent, a named owner and scoped, short-lived credentials.

Corma gives IT visibility into shadow AI and the AI tools in use, with their costs, and brings those applications into the same access reviews and deprovisioning workflows as the rest of the SaaS stack, so a departure also removes the person's accounts in those tools.

Related concepts

FAQ

Is an AI agent a non-human identity?

Yes. AI agents are one kind of non-human identity, alongside service accounts, API keys and workload identities. They stand apart because they decide their own actions.

Should an AI agent use an employee's credentials?

No, not a shared password or a personal API key. If the agent must act for a user, it should use delegated, scoped tokens, so its actions stay limited and traceable to both the agent and the user.

Who should own an AI agent identity?

A named person, usually the business owner of the use case, who answers for the agent's purpose, access and retirement. IT or security sets the rules and reviews them.

How do you offboard an AI agent?

Revoke its tokens and keys, delete its OAuth client or workload identity and remove its accounts in connected applications. Do the same when its owner leaves.

Request a demo