Agent Autonomy Preflight

A pre-flight checklist and monitoring procedure for agents that run without a watcher. Teaches the seven failure modes that only appear when nobody is looking, each one a real observed incident: state stored rather than derived so dead agents looked alive; a retry loop that reached 111 requeues and provoked the rate limit it was surviving; a failed tool returning an empty list that the agent read as "no results"; a dashboard reporting 8,762 active agents of which 82.7% was its own monitoring; four authorizations booked as revenue that never settled; a missing credential surfacing as a throttling error; and unbounded overnight spend. Includes runnable staleness-check and spend-cap code.