I had seven AI agents on cron. Two of them stopped running on day one. I noticed on day eighteen.
That sentence is the whole article, and it is also the kind of sentence I would have argued against if someone else had said it on a podcast. "Surely you would notice. You have tracing. You have dashboards. You have a Telegram channel that lights up every time anything moves." Yes. I had all of those. The two dead agents still slipped under all of them, because every monitoring layer I had was built to watch processes that ran. Two of mine were not running.
This is the eighteen-day log: what the seven agents were, how two of them broke silently on day one, why my tracing was the wrong shape for the failure, and the small exit-code contract I now bolt onto every scheduled CLI agent.
The seven agents and a setup that looked fine
I run two content domains and one self-evolving harness on the same box. Each domain has three agents on a daily cron at 09:00 (observer, strategist, marketer), plus a shared evolver that runs on Saturdays. That is seven processes. The cron lines all looked roughly like this:
