What agent identity is
When an AI agent does work inside a company's systems, the systems need to know who is acting. Agent identity means the agent gets its own account, like an employee badge, instead of borrowing a person's login or a shared password.
The identity usually carries three things: a unique ID, a set of permissions, and a named human or team responsible for it. Anything the agent does is recorded under that ID.
Why a separate identity helps
Most agents today run on the credentials of whoever set them up. That causes three problems.
First, permissions are too broad. The agent inherits everything its owner can touch, not just what the task needs. Second, logs are muddy. If the agent deletes a file, the record says a person did it. Third, shutting it down is awkward. You cannot revoke a person's access without locking out the person.
A separate identity fixes all three. You can apply least privilege, granting only what the job requires. The audit trail shows the agent by name. And revoking one agent does not touch anyone else.
How it works in practice
Teams usually build it from tools they already have. The agent is registered as a service account or workload identity in their identity system. It authenticates with short-lived credentials such as an access token, often issued through OAuth, instead of a long-lived password or a static API key. Secrets are kept in a vault, as described under secrets management.
Good setups add a few more things: an owner who is accountable for the agent, an expiry or periodic review so forgotten agents lose access, and a human in the loop approval step for risky actions such as payments or deletions.
Why it matters when you use AI
As more AI agents run on their own, the number of non-human accounts in an organization can outgrow the number of people. Each one is a possible way in if its credentials leak or if the agent is tricked by prompt injection into misusing what it holds.
If you deploy agents, ask four questions: does each one have its own identity, what can it reach, who owns it, and how fast can you turn it off?
An agent sandbox limits where an agent can run. Agent identity limits what it is allowed to do and keeps a record of it.