ERC-8004 is the crypto world’s bid to become the identity-and-trust layer for AI agents — the standard backed by authors from MetaMask, the Ethereum Foundation, Google and Coinbase, designed to answer the question MCP and A2A deliberately skip: who is this agent, and should you trust it? Its registry contracts went live on Ethereum mainnet in January 2026, and 10,000 agent identities were registered almost immediately.

Now there’s a peer-reviewed measurement of what actually got registered. A Scientific Data (Nature Portfolio) article in press — “A dataset of early blockchain-registered AI agents on Ethereum” by Yulin Liu (Quantum Economics AI Lab) — indexes the first 10,000 ERC-8004 agents end to end: identity records, mint economics, ownership transfers, off-chain metadata, declared services, and every reputation feedback record, all pinned to a single Ethereum block for reproducibility. The dataset is public and CC0 on Harvard Dataverse.

It is the first hard look inside the registry — and the registry is largely empty.

10,000 agents in 7.7 hours

All 10,000 agents were minted on January 29, 2026, within a 7.7-hour window — the day the registry went live. They arrived in 3,028 transactions, averaging 3.3 agents per transaction with a maximum of 182, which the paper reads as operator-driven batch minting rather than organic adoption.

Ownership is concentrated to match: the 10,000 agents belong to just 394 wallets. The single largest owner holds 779 agents, and the top ten owners hold 51.4% of everything (a Gini coefficient of 0.863). This was a land grab, not a groundswell.

72 agents have machine-readable capabilities. Out of 10,000.

ERC-8004’s design puts the interesting part — what an agent can actually do — in an off-chain metadata document pointed to by an on-chain token URI. The paper’s resolution pipeline found:

  • 7,856 agents (78.6%) never set a token URI at all. They are bare mints: an ID and an owner, nothing else.
  • Of the 2,144 that did set one, only 72 resolved to valid, parseable JSON. The rest returned HTML landing pages, error pages, empty bodies, or timed out.

That is 0.7% of registered agents with a real capability declaration. From those 72 agents the authors extracted 112 declared service endpoints — mostly generic web links, followed by A2A and OASF, with only a small number of MCP and x402 endpoints. Notably for anyone in commerce: none of the 112 declared services is a retail or merchant-facing endpoint.

The reputation layer is monitoring bots, not markets

ERC-8004’s second pillar is a Reputation Registry where counterparties file feedback on agents. The observed activity: 980 feedback records across 628 agents, from 197 client addresses. But:

  • A single automated issuer filed 645 of the 980 records (65.8%), almost entirely tagged liveness and trust — an uptime checker, not a counterparty.
  • The top three issuers account for 71.4% of all feedback.
  • Only 13 issuers ever rated more than one agent.

The authors’ own conclusion: early reputation activity is “dominated by a small number of automated monitoring/liveness issuers rather than broad organic peer feedback.” Because the registry is permissionless — any address can rate any agent — they ship a per-issuer screening table so researchers can filter for Sybil and self-referential patterns before trusting the numbers.

And the third pillar, the Validation Registry — the part meant to hold cryptographic proof that an agent’s work was verified — is, in the paper’s words, “not yet meaningfully populated on-chain.” It contains so little that the dataset excludes it entirely.

Why this matters for merchants

When we talk to store owners about agentic commerce, a version of this question keeps coming up: won’t there be some registry that tells stores which AI agents to trust, and tells agents which stores to trust? ERC-8004 is the most credible crypto-native candidate for that layer, and this paper is the first independent, peer-reviewed evidence of where it actually stands a few months after launch:

  1. The buyer-agent trust layer is not coming to save you soon. 0.7% real capability declarations, no organic reputation, an empty validation layer, and zero commerce endpoints. The agents actually buying from stores today — ChatGPT’s checkout, Google AI Mode, Copilot — identify themselves through user-agent strings, Web Bot Auth signatures and platform programs, not on-chain registries.
  2. Trust signals will be composable facts, not one score. The feedback records that do exist carry tags — liveness, reachability, quality — rather than a single universal rating. That matches what we see across the whole stack: machine trust is assembled from verifiable, specific claims.
  3. Registration is not readiness. 10,000 registered identities, 72 with substance, is the on-chain version of a pattern merchants should recognize: being listed is not the same as being usable. A directory entry — on-chain or off — does nothing if the endpoint behind it returns an error page.

One caveat in fairness to the standard: the snapshot ends at block 24,839,925 (roughly February 2026), and the paper’s pipeline is open-source and re-runnable, so later snapshots may look different. The 72 metadata-bearing agents are also self-selected early publishers, not a random sample. But as of the only measured moment we have, the “trustless agents” trust layer is a registry of empty shells being pinged by uptime bots.

FAQ

Does my store need to register anything with ERC-8004?

No. ERC-8004 registers agents, not merchants, and nothing in today’s agentic checkout flows (ACP, UCP, or assistant crawlers) reads it. The trust signals that affect whether an agent can buy from you live on your own site: structured data, feeds, bot access policy, and checkout reachability.

Is ERC-8004 dead, then?

Registered in a day and abandoned is a common fate for new on-chain standards, but the design isn’t refuted — the paper measured adoption, not soundness. The signal to watch is the Validation Registry: if verifiable attestations of agent work start landing there, the trust layer becomes real. As of this snapshot it is empty.

Who is behind this data?

An academic dataset paper (in press at Scientific Data, a Nature Portfolio journal) by Yulin Liu of Quantum Economics AI Lab, built by querying the ERC-8004 contracts on Ethereum mainnet directly, with every state read pinned to one block so the results are independently reproducible. The full dataset is public domain.

Sources

An on-chain registry can’t tell an agent whether it can buy from your store — your site does that, every time an agent visits. AgentReady’s scan measures exactly that chain: can an agent find, understand and buy from you.