Agentic identity

Agentic identity is the discipline of establishing who an AI agent acts for, what it is currently permitted to do, and how its authority ends. It differs from classic machine identity because agents change behavior over time, delegate to other agents, and act with human authority, so the interesting questions are about the delegation and its revocability, not the credential.

The question that classic identity does not answer is scope over time. A service account is provisioned once and its permissions are reviewed quarterly; an agent may be handed a task at ten in the morning that requires one capability and a task at noon that requires another, and the authority it accumulates between them is the exposure.

The other difference is attribution. When an agent acts, something has to record which human's authority it was spending, or an audit cannot answer who approved the action.

Scope over time is the question classic identity management has no vocabulary for. A service account is provisioned once and reviewed on a cycle; an agent may need one capability at ten in the morning and a different one at noon, and if both grants persist it ends the day holding the union of everything it has ever needed. The accumulation is the exposure, and nothing in a quarterly review catches it.

Attribution is the second gap. When an agent acts, the record has to say whose authority it was spending, or an audit cannot answer who approved the transaction, and an incident cannot determine what else that authority reached. A delegation chain is the structure that answers both in one traversal.

Termination is the third, and it is the one most likely to be left implicit. An agent's authority should have an end condition that does not depend on someone remembering to revoke it, which is the argument for just-in-time capability tokens over durable grants: the authority expires with the task rather than persisting until it is noticed.