Guardrails, review rules, kill switches, and monitoring routines for AI agents that take real actions.