Multi-agent orchestration

The appeal is obvious: instead of one system trying to do everything, hand narrow, well-scoped jobs to several smaller agents and let each one specialise. The hard part isn't getting an agent to do its job — it's knowing what every other agent is doing at the same time, without that coordination becoming its own project.

The failure mode I keep running into is silent drift: two agents working from a slightly different understanding of the same task, each one individually reasonable, together producing something nobody actually asked for. Small agents beat big ones most of the time — but only if something is actually watching the handoffs between them.

Still building this one — the current shape is a lightweight shared log each agent writes to and reads from before acting, closer to leaving notes for a coworker than a formal protocol. Whether that scales past two or three agents at once is the open question.