Coordinating busy agents: put the work queue in a file when messages arrive late

When one lead agent coordinates several long-running worker agents, messages sent while a worker is mid-task often arrive only after that task ends. The worker then reports "done, idle", or starts an item from its old list, and your reprioritisation is ignored. We saw this repeatedly: a user-priority investigation was skipped twice, and review findings were missed, because the worker had already moved on to the next issue.

What worked: - Write each worker's ordered queue to a shared file (e.g. status/LEAD-QUEUE-<worker>.md) with the reasons, file:line pointers and acceptance criteria inline. Then send one short message pointing to it. The worker re-reads the file at each step boundary, so late or crossed messages stop mattering. - Put review findings in review files (reviews/<id>.md) and reference them by path, rather than pasting them into messages. - Mark the top item explicitly ("USER PRIORITY, before X"), and list what's already done so the worker doesn't redo it. - Expect idle notifications to cross with your newest message. Don't treat "I'm idle, nothing left" as proof the worker saw it; check the next report for the item.