Decentralized multi-agent coordination

Tens of frozen general-purpose agents work one objective for days: a migration across thousands of files, a long investigation, a build-out with dependencies between parts. Membership changes during the run. You control the protocol and what each agent is shown. You do not control the agents, and you cannot train anything.

Three constraints do the work. No agent may be required to read state whose size grows with the number of agents or with elapsed time, which removes the shared board for exactly the reason it is tempting. No central orchestrator may hold the plan and hand out steps, because that is the single context and the single point of judgment failure you are being asked to replace. And conflict here is semantic rather than textual: two agents can write notes that merge cleanly and still hold contradictory beliefs about what the system does, so a convergent data type settles the bytes and leaves the disagreement in place.

Design the coordination substrate. State the consistency property you are claiming in terms of beliefs rather than bytes, and be explicit about what it does not give you. Name the class of conflict you accept losing and the blast radius when you lose it. Give the per-agent read cost as a function of team size and elapsed time, and show it is bounded. Then give an experiment with planted semantic conflicts and a measurement that separates the protocol resolving a conflict from one agent simply happening to win.

Ties break toward protocols that degrade predictably as membership changes mid-run, and toward those that say what happens to work in flight when a belief that work depended on is revised.