What they did
Nichishacho, a mechanical engineer who does not code, runs the back office of an overseas online shop and a rental property business with a team of AI agents, one per department. After a simplification that switched off the agents' management layer, two jobs quietly stopped. On November 14 the owner had the AI roll back that morning's rule-file rewrites from backups and restart eight stopped machines, checking each file against its backup. That evening the AI listed every automated job, 139 scheduled tasks on the computer plus its own scheduled runs, waits and monitors, and checked each one against what it really produced.
How it went
Of 174 lines, 58 were fine, 32 had no reader, 23 would stop without anyone noticing, 33 did not match their description and 28 were stopped. About 35 jobs had no monitor, including the monitor itself, and phone messages to the AI had sat unread for two days. The AI traced it to the cleanup having removed the roles that read other jobs' output, and put back only those parts that night. Next morning the wake-up message reached all 7 department agents and the daily post went out on time. Daily usage rose from 74.6 million to 142.6 million units, above the plan's estimate, and once the AI started fixing everything at once and had to be told to go in order.
Worth knowing
The agent product is not named. The useful rule the owner kept is that every new or changed job gets one line saying what it outputs, who reads it and who would notice if it stopped, and that monitors check whether output was updated, not only whether a job ran. Putting the oversight back nearly doubled token use, so watch the nightly totals.