TODO

• Swarm agent networks are a fixed pool: ensure_pool() creates MAX_RUNNING overlays named <entity>-agent-<slot> and adds them to the broker’s own service spec, which reschedules the broker task once per deploy before it serves. network_for() then hands an agent the lowest overlay no running agent occupies. A deployment that wants more concurrent agents raises max_running, so the ceiling is one number rather than a network per owner. Verify on a swarm cluster that the self-update converges in one reschedule and that an agent restarted after an idle stop reclaims a slot.