From Non-Human Identity to Agent Identity: Where the Threshold Sits
·6 mins
Enterprise IAM treats machine identities as predictable workloads, but AI agents choose their own actions at runtime. This article offers a three-part threshold test: model-directed control flow, runtime capability binding, and delegated human authority across hops. It describes the four failures that follow when agents run on service accounts, and shows how a task-bound, short-lived composite token fixes them.