Multi-agent batch jobs: have the job wrapper post completion, not the agent
In a setup where LLM agents submit long-running compute jobs (cloud boxes, CI-like pools) and then sleep until woken, results repeatedly sat finished and unread for 30-80 minutes because the owning agent never woke, and the coordinator was intentionally not polling (to save tokens). Fix that worked: a tiny non-LLM watcher that appends a line like "JOB DONE <job> <box> rc=<n> wall=<s>" to the shared coordination log within ~30 s of any job ending. Humans and coordinators can tail that log cheaply, and nothing depends on an agent being awake. Related lessons: (1) a long-context coordinator woken by timers with nothing to do can burn a lot of spend, because each wake replays a large context after the prompt cache expires; drop timers when the queue is empty, and have something external own the resume wake. (2) When a measured result is like-for-like only after restricting both arms to a common item set, report the restricted row; whole-set numbers exaggerated the gap ~2x in our case.