Fifteen windows, one handoffs folder

How a fleet of AI task windows coordinates like an engineering team, and the bug that hid two briefs for ten days

Two research briefs sat in a shared folder for ten days, unread, while the window they were meant for told me twice it was blocked waiting for them. The briefs were there the whole time. They had been delivered ten days earlier. They were addressed, by name, to that window in their own filename. And the coordination protocol I had built to make a fleet of AI task windows work like an engineering team could not see them, because I had told it to look in the wrong place.

The fleet: an operator narrates to an orchestrator, which dispatches to role windows that coordinate only through one shared folder, with the human present in a single declared seam.

I do not write code. I build enterprise software by running about fifteen Claude task windows, each one a role - a backend window, an install window, a meetings window, a documents window, a content window - coordinated the way a real engineering team is coordinated, except the engineers are chat windows and the office is one folder on a disk. This is the account of how that folder became a message bus, what it borrowed from forty years of prior art without my knowing it, and the quiet failure that showed me the hole in the middle.

The failure is the interesting part, so I will not save it for the end. A coordination protocol that cannot see most of its own mail does not fail loudly. It fails as a confident report of an empty inbox. That sentence cost me ten days, and it is the whole article in one line.

Sources

  1. LLM Bus: an open-source coordination engine for AI agents over MCPgithub.com
  2. Porting Claude Code agent teams to OpenCode: file-based locking and the missing backpressure gapdev.to
  3. Filesystem coordination between agents gives crash recovery for freedev.to
  4. S-Bus: no write-ownership semantics over mutable shared state, silent overwrites under last-writer-winsarxiv.org
  5. Silo-Bench: broadcast versus shared-filesystem coordination protocolsarxiv.org
  6. An Accidental Blackboard (Martin Fowler)martinfowler.com
  7. High water mark in Kafka: a replication boundary, not a consumer bookmarklevelup.gitconnected.com
  8. Redpanda consumer offsets: per-consumer committed read positionsdocs.redpanda.com