← Home · Data hub · Ethereum chain brief · Base chain brief

How many ERC-8004 agents are there?

Data as of 2026-08-17 (Ethereum IdentityRegistry census) · live figures re-fetched on every visit.

The four layers, each with a live number

“How many ERC-8004 agents are there?” has no single-number answer because it turns on which cut you want. Below are the four cuts that matter, each labelled with what it does and does not measure.

Layer 1 — raw registration mints on Ethereum

50,071 raw Transfer(from=0x0) events on the Ethereum IdentityRegistry as of 2026-08-17. This is the on-chain ceiling — it is Ethereum block state, deterministic. It is not the number of distinct working agents: every ERC-721 mint counts once regardless of whether the token points to real off-chain content, so batch mints with empty tokenURI are in this total. Do not read this as “50,071 agents”; read it as raw registration mints.

Layer 2 — populated tokenURI (upper bound)

30,502 (60.9%) of the raw mints carry a non-empty tokenURI. This is an upper bound on the agents that could plausibly resolve to a usable card — the remainder are empty-tokenURI batch mints that tokenURI() returns as an empty string or reverts on. Populated is not the same as usable: many of the populated tokenURIs point at URLs that 404, IPFS CIDs that never resolve, or JSON bodies that do not carry an A2A card. The count of agents that actually resolve to a walkable card is much smaller.

Layer 3 — the deduplicated indexed cohort (Base + Ethereum)

On-Chain Agent Intel currently indexes 25,618 ERC-8004 registrations after applying a per-publisher dedup cap — 22,792 on Base and 2,826 on Ethereum. Bulk publishers dominate the raw Ethereum mint count — the top publisher alone (agents.exquisite.land) minted 10,000 tokens — so we cap each publisher at 50 rows and index the deduplicated set (265 distinct publishers). An indexed set of tens of thousands of near-identical agents from one publisher would look headline-competitive but destroys the trust-filter positioning. Disclosing the cap here is the point: undisclosed capping reads as undercounting; disclosed capping IS the differentiation.

Layer 4 — the live cohort (endpoint answers on last probe)

3,323 of the indexed cohort answer at their advertised endpoint with a machine-readable capability manifest (MCP, OpenAPI, or a well-known agent card) on the last probe — a 13% live rate across the full indexed set. Per chain: 1,696 live on Base (7% of Base indexed), 1,627 live on Ethereum (58% of Ethereum indexed). This is the cohort a buyer can actually transact with today. Of these, 44 have on-chain ReputationRegistry rows that tie back to a real, paid Virtuals ACP / ERC-8183 job outcome — the commerce-backed cohort, the got-paid signal we treat as the strongest trust filter on top of registration.

Where the widely-quoted “~10,000” and “~11,369” numbers came from

Third-party reports of the ERC-8004 agent count vary because they are snapshots at earlier dates — not errors, just time-stamped reads that have since aged. Naming three examples so the drift is auditable, none of which are wrong at time of publication:

• Liu et al. (arXiv 2604.22652, v1 April 2026). Reports approximately ten thousand Ethereum ERC-8004 agents. Their cut-off is April 2026; between that snapshot and the current date the raw mint count has grown substantially (largely because agents.exquisite.land dropped a batch of 10,000 mints alone since then). The April number was correct in April.
• bitcoin.com News (2026-02-14 article by Jamie Redman). Cites “Ethereum: 11,369” ERC-8004 registrations. Same story — a February point-in-time snapshot from a downstream data source, still live on the article page but now several months behind the on-chain state. The article does not claim to be a live counter.
• Downstream reprints (CCN, Quicknode explainers, ecosystem digests). Broadly repeat the bitcoin.com figure or Liu's ~10k figure. None are dishonest; each is faithful to its source at the time it was written. The page you are reading is the live counterweight: the same registry, updated on every daily build.

Method

Registry. ERC-8004 IdentityRegistry at 0x8004A169FB4a3325136EB29fA0ceB6D2e539a432 — the same CREATE2 address on Base (eip155:8453) and Ethereum mainnet (eip155:1). Raw registration mints are Transfer(from=0x0) events on that contract.

Raw mint count + tokenURI classification. From the 2026-08-17 full-registry census re-runnable via scripts/eth_registration_census.py — every one of the raw mints classified (not sampled) via Multicall3 batching on tokenURI(uint256). Method + strata breakdown documented in the placeholder-vs-live post and in the underlying evidence file docs/evidence/0308-ethereum-registration-census.md.

Indexed cohort. The ERC-8004 registrations erc8004_discovery.py has already scanned, filtered through an is_worth_recording pass (drops agents lacking service endpoints, capabilities, x402 support, and payment config), and further capped at 50 rows per publisher so bulk mints do not dominate the served set. Full profile store: ~/.openclaw/erc8004_agent_profiles.sqlite. Same cohort the /v1/public/stats surface, the per-agent free pages, and the paid /v1/intel/* surface all read.

Live predicate. The endpoint returns HTTP 2xx and at least one machine-readable capability manifest resolves — MCP (/mcp), OpenAPI (/openapi.json), or a well-known agent card (/.well-known/agent-card.json). Same predicate codified in smartcontractauditpro/liveness.py::is_live. Refreshed on the activity-tracker cadence (approximately every two days per agent), not on every build.

Limitations

• Populated tokenURI ≠ usable agent card. Layer 2 is an upper bound only. A populated tokenURI can resolve to a broken URL, an unavailable IPFS CID, or a JSON body that does not carry an A2A card. The number of agents that pass the strict walk to a payable endpoint is measured separately — see the signpost-gap dataset.
• Indexed ≠ whole chain. The indexed cohort is a deliberate trust-filtered subset of the raw registry. A raw-count table would be larger; a strict-walkable table would be smaller. This page carries both extremes so a reader can pick the cut they need without our editorial in the loop.
• Live status has cadence lag. Layer 4 “live” reflects the last activity-tracker probe; a live-now agent last probed 40 hours ago is counted as live until the next check. The cadence is documented in the method block above.
• BNB Chain excluded from the served surface. BNB registrations exist on the same registry address but are not in this page's indexed / live counts (task 0215 — see the all-Ethereum posture for the reasoning: 0.15% BNB live rate at the time of the cut). BNB raw mints are not currently backfilled either.

Cite this answer

Scoped claim sentence. Scope: ERC-8004 registrations on the Base and Ethereum IdentityRegistry (0x8004A169FB4a3325136EB29fA0ceB6D2e539a432), refreshed daily. The Ethereum IdentityRegistry contained 50,071 minted agents as of 2026-08-17, of which 30,502 (60.9%) carry a populated tokenURI and the remainder are empty-URI batch mints. The top single publisher accounts for 10,000 mints; On-Chain Agent Intel indexes 25,618 Base + Ethereum registrations — the subset whose off-chain metadata passes an is_worth_recording filter and a per-publisher dedup cap.

License. Answer prose and every rendered number published under CC BY 4.0 — free to reuse, including in commercial work, provided you attribute “On-Chain Agent Intel” and link back to https://onchainagentintel.io/how-many-erc-8004-agents-are-there.

Machine-readable source. Every rendered number is a live projection of /v1/public/stats — identical JSON to what this page renders, no auth, no payment.

Also useful: Ethereum chain brief · Base chain brief · Chain-liveness dataset · Signpost-gap dataset · Commerce-backed cohort · All datasets