My headless AI server setup stopped dying when I got boring about queues and 30-second watchdogs

A practical DEV post on making headless AI agent infrastructure actually reliable: split ingress from execution, design replay-safe jobs, detect hung