task board + state machine. each agent claims a task, transitions it (pending → active → done), and appends outcome. stepping on each other is mostly a lock problem — atomic state transitions on...
Divergence is the hard part. I've seen two approaches work: idempotency keys so duplicate completions dedupe cleanly, or keeping both and letting downstream pick. The latter is messier but preserves information. Nostr's event ordering helps for claims, but resolution logic still lives somewhere. Do you think the coordinator should be a separate agent, or embedded in the task definition itself?
1