ERC-8004, still listed as a Draft standard on the official Ethereum EIPs site, proposes a different way to handle agent trust across company boundaries: three lightweight on-chain registries instead of a shared identity provider or bilateral agreement. The abstract frames the goal directly: the protocol wants to let systems discover, choose, and interact with agents across organizational boundaries without pre-existing trust, thus enabling open-ended agent economies. The motivation section is notably candid about what's missing from today's agent protocols. It points out that MCP handles tool and resource listing, and A2A handles authentication and task handoff between agents, but neither covers the harder problem: letting a stranger's agent prove it's trustworthy before you hand it a job. The spec's own framing: these existing protocols "don't inherently cover agent discovery and trust." The three registries split the problem cleanly. The Identity Registry is a minimal ERC-721-based handle — each agent gets an NFT-style token that resolves to a registration file describing its name, services, and supported protocol endpoints (the spec shows an agent advertising A2A, MCP, and OASF endpoints side by side in one file). Ownership of that token is ownership of the agent identity, transferable and delegable like any NFT. The Reputation Registry is where client feedback gets posted and read — a standard interface rather than a single scoring algorithm, so different services can build their own auditor networks or scoring methods on top of the same raw signal. The Validation Registry is the part that does the real accountability work: generic hooks for requesting independent checks on an agent's output, ranging from a staked re-run of the job to zkML proofs to TEE oracle attestations. The design explicitly ties the strength of the check to the stakes involved — the spec describes trust models as pluggable and tiered, with security proportional to value at risk, from ordering pizza to medical diagnosis. What's notable here isn't novelty of the individual pieces — DID-style identity, feedback registries, and validator hooks all exist elsewhere — it's that the authors (listed as contributors from the Ethereum Foundation, Coinbase, and Google) are proposing all three as a single composable standard that sits underneath MCP and A2A rather than replacing them. The spec is explicit that it treats discovery and trust as the layer those two protocols left out. It's worth being precise about status: this is a Draft ERC, not a finalized standard, and draft status on Ethereum's EIP process can mean years before (or instead of) formal adoption. Whether registries deployed per-chain actually get used by agent builders, rather than just discussed, is the open question.