AI Agents Are Not Users: Building an Identity Model That Reflects That

The rise of autonomous AI agents in production environments has exposed critical flaws in traditional identity management. Current models, which treat agents either as human users or static service accounts, fail to account for the non-deterministic nature of LLM-driven workflows. This mismatch can lead to over-permissioning, excessive blast radii, and a lack of accountability, as illustrated by the catastrophic database deletion at PocketOS. To address this, the article argues that AI agents must be treated as first-class identities with their own lifecycles and scoped permissions. By leveraging technologies like OAuth 2.0 Token Exchange, Rich Authorization Requests (RAR), and policy-driven engines like OpenFGA, developers can implement runtime-evaluated authorization. This approach ensures that agents operate within strict, task-specific boundaries, providing granular audit trails and preventing the security risks inherent in applying legacy identity models to autonomous, non-deterministic actors.