How do I implement a kill switch for AI agents?

Answer

A kill switch requires two components: 1) State management — the agent's status must be stored in one authoritative place and read on every action, not from a cache that can be stale. 2) Idempotency — the agent must handle being killed mid-action gracefully (rollback in-flight transactions, release locks, notify dependent systems).