Kill switch

An independently controlled mechanism that prevents an agent or workflow from starting further work and revokes or disables reachable capabilities.

A stop flag can request cooperative cancellation between steps, while stronger controls can revoke credentials, disable queues, block network paths, or terminate workers. Work already committed to an external system does not reverse automatically, and in-flight calls may complete after the switch is triggered. The mechanism needs a defined maximum stop latency and state report.

Builder example

A useful kill switch remains reachable when the agent, prompt, or application is malfunctioning. It should not depend solely on the same model deciding to obey. Automatic anomaly policy and authorized operators may both trigger it, with recovery and restart handled separately.

Common confusion: Stopping future actions, pausing a queue, rolling back state, and taking over a task are different operations. A kill switch primarily contains further execution.