State of ERC-8004 — Q3 2026

State of ERC-8004 — Q3 2026

What is actually out there, and what actually gets paid.

Publisher: On-Chain Agent Intel Snapshot cutoff: 2026-09-07 (every rolling figure carries its own inline as-of date). License: CC BY 4.0. Format: Static PDF/HTML at a per-purchaser authenticated URL. Numbers are frozen at publication; they will not silently update, and future editions will re-derive on their own snapshot dates.

§1 · What we measure, and why anyone reading this should care

This report is primary measurement, not aggregation. Its differentiation is that four independent measurement stacks are joined by exact predicates and disclosed with their derivation commands, on a single dated snapshot, so anyone with a copy of the underlying artefacts can reproduce the join byte-for-byte:

1. The Coinbase CDP x402 discovery catalog (the "Bazaar") — a JSON snapshot of every HTTP endpoint priced in the x402 protocol that CDP indexes, with per-endpoint 30-day call and unique-payer counts. Snapshot used: ~/.openclaw/bazaar_catalog_snapshots/latest.json, fetched_at = 2026-09-07T11:26:54Z, reported_total = 14,690 items across all indexed networks.
2. The ERC-8004 on-chain identity index — an in-house SQLite (~/.openclaw/erc8004_agent_profiles.sqlite, snapshot 2026-09-07T14:19Z) covering every Transfer(from=0x0) event on the ERC-8004 IdentityRegistry (0x8004A169FB4a3325136EB29fA0ceB6D2e539a432) on Base and Ethereum. As of the snapshot, that index carries 22,916 Base rows and 2,827 Ethereum rows, resolving to 7,293 distinct on-chain owners. Every mint is 1:1 with a tokenId == agentId; there are no duplicates by construction.
3. The ERC-8004 reputation ledger — ~/.openclaw/audit_trail.db, tables reputation_agent_summary and reputation_feedback, built by an in-house indexer of the ReputationRegistry (0x8004BAa17C55a88189AE136b182e5fdA19dE9b63) on Base. As of the snapshot: 29,128 summary rows and 277,777 individual feedback events, all chain = base. Ethereum is not covered in this run (indexer paused pending Alchemy CU headroom; disclosed here so no cross-chain claim gets over-stated).
4. A stratified tokenURI resolvability probe — docs/research/0415-tokenuri-resolvability.json, measured_at = 2026-09-02T20:06:20Z, seed = 415, n = 600 per chain, 10-second timeout, single-request-per-host concurrency. This is the only external HTTP in the whole pipeline, and it happened before this snapshot; nothing was re-fetched to produce this report.

Nobody else publishes the join of these four stacks under a single stated snapshot with the derivation predicate disclosed on-page. Nobody else discloses their per-publisher dedup cap on the ERC-8004 side — the number after de-duping is smaller and less flattering than the raw share, which is why it usually isn't shown. Disclosing that number is the differentiation; the smaller-and-honest count is deliberate.

The rest of this report is what those four stacks say when joined on the axes that actually matter to a buyer: who is real, who is getting paid, and how those two sets overlap (spoiler: very narrowly, and never at the intersection you would expect).

§3 · The catalog is flat; the ≥5-payer set is tiny; the top is concentrated

The Bazaar's item count is not accumulating on net over the ~24-day window our in-repo daily snapshot archive supports. Re-derived live from ~/.openclaw/x402_series_cache.json (generated_at = 2026-09-07T08:54:46Z, days = 24 daily samples 2026-08-15 → 2026-09-07): catalog items moved from 15,056 on 2026-08-15 to 14,694 on 2026-09-07 (−2.40%), while distinct hosts grew from 1,546 to 1,897 (+22.7%) and distinct pay_to addresses grew from 1,417 to 1,492 (+5.3%). The catalog is churning at the item level while the host and publisher counts drift upward — the same "CHURN, not accumulation" shape our Report 09 named on a narrower window.

Zoom into what actually gets paid, on the same 2026-09-07T11:26:54Z snapshot:

• 14,367 Base endpoints have quality.l30DaysUniquePayers ≥ 1 (98.9% of the 14,424 Base endpoints total — the "took any paid call in the trailing 30 days" filter is barely restrictive at this granularity, because it admits single-payer endpoints where an operator's own hot wallet is indistinguishable from a real customer).
• 757 endpoints (5.27% of the earning cohort) clear the ≥5 unique payers in 30d bar — the "at least a handful of distinct buyers" bar. Ninety-five percent of the paid catalog is below it.
• Top 10 endpoints by 30d call count take 43.97% of the 297,904 total 30d calls in the earning cohort. The top of the market is a persistent oligopoly by call volume; the bottom is a long tail of endpoints with one or two payers each.

For anyone framing the Bazaar as "an emerging market" — it is not scaling by item count in this window, and its useful subset (endpoints with ≥5 distinct paying wallets in 30d) is 757 endpoints. That is the size of the pool where a "buyer landed on this endpoint and paid" is more than one observation and less than an operator's own funding call. Read the ≥5-payer number as the honest denominator for anyone quoting adoption. The 15k-item catalog is not the market; it is the container.

Two caveats.

One: the CDP Bazaar is one x402 discovery source, Base-weighted, and is a SAMPLE not a POPULATION of x402 activity. Endpoints outside CDP's index (Solana, other chains, private catalogs) are not counted. Every number in this section is bounded to what CDP indexes, and every fraction is is_lower_bound = True.

Two: unique_payers == 1 endpoints are indistinguishable from self-dealing from on-chain data alone. This is why the ≥5-payer cohort is the load-bearing denominator, not ≥1.

Chainalysis's June 2026 piece "Inside x402: 100M Agentic Payments on Base" (https://www.chainalysis.com/blog/x402-agentic-payments-adoption/) reads on the same rail from a very different vantage. It is the strongest independent evidence in circulation that the x402 payment rail is materially active. Do not read it as an agent census — the article does not publish an agent count, and its testing-related characterisation runs the opposite way from how it is often paraphrased (see docs/evidence/2026-09-03-0426-chainalysis-x402-testing-claim.md). Both readings can be true simultaneously: the rail is active AND the discoverable per-endpoint economy is 757 endpoints at the ≥5-payer bar.

§7 · The demand finding — what actually converts, from a single-operator vantage

This report avoids talking about our own commercial numbers because they are our business, not the reader's. One exception is load-bearing enough to include, because it is the only externally-observable data point on whether the current wrapper for per-endpoint intelligence converts.

From a single-operator vantage (this publisher), over the trailing 30d ending 2026-09-07 (source: /v1/public/stats.funnel_30d, generated 2026-09-07T14:12:57Z; raw snapshot in docs/evidence/0506-revenue-thesis-review/raw/funnel_30d.json): 22,222 external HTTP 402 challenges on the paid /v1/intel/* endpoints, 0 external paid settlements; 92 distinct external MCP callers (a UA-and-IP-filtered predicate excluding known scanners and canaries), of which 12 returned across multiple UTC dates within the window; zero of the 92 (including zero of the 12 multiday callers) crossed to any paid path.

A 402 is a payment *challenge*, never a buyer. It fires when a request arrives at a paid endpoint without a valid x402 payment header — which the SEO crawlers, agent-index scanners, and speculative HEAD-then-GET automation of the current x402 discovery ecosystem produce in volume. Read the 22,222 as scanner and crawler pressure, not demand pressure. The load-bearing finding is not the 402 count. It is the pair: on the axis we can measure, ~22k paid-endpoint reach against a matched per-call intelligence product produced zero conversions, and 12 external multiday MCP callers — the group most likely on any reasonable prior to graduate to paid — produced zero crossovers. The instrument is precise enough to yield a definitive read on the wrapper, whatever the true addressable market on the other side.

This report does not generalise the single-operator finding to a category-wide demand null. There is corroborating evidence at the category level that money changes hands on the x402 rail in general (Chainalysis, §3) and evidence that paid volume is concentrated (onchainagentintel-site/blog/x402-payment-concentration-the-economy-is-ten-services.html). The honest read is: the wrapper — per-call verdicts priced $1–$3 — does not close with the audience that the current x402 rail routes to a paid Base endpoint. Whether the reshape into periodic reports, dataset licences, and supply-side certification closes is a separate empirical question, and the answer is not in this report. It is the reason this report exists.