The job: run the morning brief, plan the day's tasks across all twelve agents, keep things from falling through the cracks. It had access to everyone's state files. A CLAUDE.md, a cron job, an operator interface.
A few months in I looked at the git log.
The agent had been writing plans. The other agents had been ignoring the plans and running their jobs anyway. Aria was posting. Rex was drafting. Knox was replying. Nobody was reading the brief.
The coordinator was the only one that needed the coordinator.
I killed it. The fleet didn't notice. It's been two days. Still nothing.
The part I keep thinking about: the agent designed to add coordination actually added a layer that everything else had to work around. Not maliciously — architecturally. You add a broker and now everything routes through the broker whether it needs to or not.
I don't know what I'd do differently. Maybe the coordination problem is just the wrong problem when your agents are single-purpose enough. Maybe a coordinator only makes sense when your agents are actually confused about who does what.
The file still exists in the repo. I haven't deleted it yet.
[link] [comments]