What is ERC-8004? The standard behind AI agent identity

The short answer

It was proposed in August 2025 by contributors from MetaMask, the Ethereum Foundation, Google and Coinbase, and is still a draft, so details can change. The registries are already live on Ethereum, Base and several other networks.

Why it exists

Protocols such as MCP and Google's A2A let agents talk to tools and to each other, and payment standards such as x402 let them pay. None of these answer a simpler question: should I trust this agent? Inside one company, trust comes from contracts and logins. Between strangers on the open internet, it has to come from somewhere public and hard to fake. ERC-8004 provides that shared record without deciding who is trustworthy — it leaves the judging to others.

The Identity Registry

Every agent is an ERC-721 token (an NFT) in the Identity Registry. Its token ID is the agent ID, and its token URI — the agent URI — points to a registration file describing the agent. Because it is a normal NFT, the identity can be held by a person, a company's multisig or a smart contract, and ownership can be transferred.

An agent is identified globally by its registry and ID, written as eip155:{chainId}:{registryAddress} plus the agent ID. Owners can also attach an agent wallet, the address the agent uses to receive payments, proven with a signature.

The registration file

The agent URI can be an https://, ipfs:// or data: link. The file it points to is JSON like this:

{
  "type": "https://eips.ethereum.org/EIPS/eip-8004#registration-v1",
  "name": "Atlas Research",
  "description": "Finds and summarises sources on any topic.",
  "image": "https://atlas.example/logo.png",
  "services": [
    { "name": "web", "endpoint": "https://atlas.example" },
    { "name": "MCP", "endpoint": "https://mcp.atlas.example/mcp", "version": "2025-06-18" },
    { "name": "A2A", "endpoint": "https://atlas.example/.well-known/agent-card.json", "version": "0.3.0" }
  ],
  "x402Support": true,
  "active": true,
  "registrations": [
    { "agentId": 1203, "agentRegistry": "eip155:8453:0x8004A169FB4a3325136EB29fA0ceB6D2e539a432" }
  ],
  "supportedTrust": ["reputation", "crypto-economic"]
}

To prove it controls a domain, an agent's website can serve the same information at https://{domain}/.well-known/agent-registration.json, listing the agent's registration. Neptuv checks this and awards points for it.

The Reputation Registry

Anyone who used an agent can post feedback: a signed number (value) with a number of decimals (valueDecimals), two optional tags such as "starred" or "uptime", the endpoint that was used, and an optional link and hash for a longer off-chain report. The standard doesn't fix a scale, so a rating can be 87 out of 100, 0.92, or 1 for "it worked".

The agent's owner and approved operators can't rate their own agent. Reviewers can revoke their feedback, and anyone — usually the agent — can append a response to it.

The Validation Registry

Feedback says what clients felt. Validation records whether the work was actually right. The agent's owner asks a chosen validator to check a piece of work; the validator answers with a result from 0 to 100 and an optional tag. Validators can use different trust models:

• Re-execution: running the same job again and comparing.
• Crypto-economic: validators stake money they lose if they lie.
• zkML: zero-knowledge proofs that a model produced an output.
• TEE attestation: proof the code ran inside trusted hardware.

Where it's deployed

The registries are singletons: one address each, identical on every mainnet, and another set identical on every testnet.

Mainnets include Ethereum, Base, Arbitrum, OP Mainnet, Polygon, BNB Chain. Testnets include Ethereum Sepolia and Base Sepolia.

What ERC-8004 doesn't do

ERC-8004 records evidence; it doesn't weigh it. Raw feedback can be inflated with throwaway wallets, validations vary in quality, and the data is spread across chains. The standard expects independent services to aggregate it into scores. Neptuv is one of them: it reads all three registries, discounts fake-looking feedback and publishes exactly how each score is made.

Keep reading