What I Learned Trying to Revoke an AI Agent Mid-Task

When an AI agent is running and something goes wrong, stopping it can be challenging. Rotating the API key affects all agents, while killing the process loses its state. This article outlines four key principles regarding the separation of identity and credentials, the need for immediate revocation, proper delegation, and the importance of audit logs in enforcement. These principles have been integrated into the open-source project Chancery.