How do I implement a kill switch for AI agents?

Answer

A kill switch requires three components:

1. State management — the agent's status must be stored in a fast-access store (database plus Redis cache) and checked on every action.
2. Propagation — when triggered, the kill switch must propagate to all running instances within milliseconds via Redis pub/sub.
3. Idempotency — the agent must handle being killed mid-action gracefully (rollback in-flight transactions, release locks, notify dependent systems). Target: under 100ms from trigger to full halt.