Analysis generated from our live index of every ERC-8004 agent we can see on Base and Ethereum. Numbers are pulled live; nothing here is hand-curated.
Two shapes at once. The Coinbase CDP Bazaar — the largest single public x402 discovery catalog
on Base — is not accumulating items on net across the ~20-day rolling-window span the in-repo
daily snapshot archive supports. The subset that has actually crossed the “≥5 distinct paying
wallets in 30 days” bar is growing much faster than the catalog: 563 → 727 endpoints
(+29.13%), with 72.82% of the earlier cohort persisting into the current window and 317 net new
entrants across the 5-payer bar. Persistence is the other half of the same shape: 10 of 10 earlier
top-10 endpoints by 30-day calls are still earning; 7 of 10 remain in the current top-10; their
share of measured calls fell 54.76% → 39.82% (−14.94 pp) as demand fanned out to new
entrants. Reference case: x402.tavily.com/search dropped from
25,073 to 1,899 calls (−92.43%) yet remains in the earning set. Two 30-day rolling windows,
~20 days apart, overlapping by roughly 10 days — not a 30-day-apart delta and no figure here
is a permanent number. SAMPLE (CDP Bazaar catalog rows), not POPULATION; a 402 is a payment
challenge, never a buyer; per-publisher dedup capped at 50, disclosed on-page. CC BY 4.0.
Flagship series update. Two months after Report 05, the ERC-8004 identity registry on Base and Ethereum
has grown, its publisher tail has hardened around a small number of big minters, and the gap between a raw
registry mint and a payable counterparty is now visible in four distinct on-chain cohorts we measure
separately: registered (a raw Transfer(from=0x0)), signposted with a
discoverable /.well-known/agent-card.json, resolvable to a usable card body,
and commerce-backed. The Ethereum registry emitted 50,071 raw registrations as of the last full block-range
census (2026-08-17) — a raw-mint total, never an agent count; 60.9% carry a populated
tokenURI, far fewer resolve to a usable agent card. A fresh 600-per-chain
resolvability sample bounds the middle layer (Base pooled usable-card 56.3%, Ethereum 71.3%, pooled 63.8%; 95%
CIs on-page, sample not census). The signpost gap is 5.93% of indexed registrations publishing a discoverable
card (20.03% of endpoint-live registrations). Per-publisher dedup capped at 50, disclosed on-page. Solana's
1,477 raw 8004-solana registration accounts from 185 distinct creators (top-10
creators = 78.67%) are reported separately as a non-ERC-8004 registry, with no resolvability claim about the
Solana rows. CC BY 4.0.
Methodology map. A buyer facing an unknown x402 endpoint has four different questions, and the public scoring
surfaces each answer a different one. Names the predicate ladder: advertised (a URL somebody published),
answers a probe (Agents Trust measures this at 138k-URL scale on their landing page; ForgeMesh's own
August census reports 1 in 4 x402 Bazaar sellers failed a probe-based sale-completion check, predicate
= envelope + reachability + paywall-fires), was actually paid (x402.fuchss.app publishes per-endpoint
30-day settlements, distinct payers, and USDC volume at payTo-wallet level,
with the caveat they publish themselves — settlement figures are wallet-level, not endpoint-specific
— and their scope statement that they probe and observe but do not verify delivery-after-payment),
counterparty is a registry-owned identity (our ERC-8004 × x402 ownership join — as of the
2026-09-01 CDP Bazaar snapshot, 14,338 earning Base x402 endpoints → 1,291 (9.0%) ERC-8004-owned
→ 26 distinct pay_tos, top publisher 67.2%, 4 with any feedback, 0 commerce-backed; per-publisher
dedup capped at 50, disclosed on-page). Nobody else joins
the paid-endpoint cohort to on-chain ERC-8004 identity. Layer 5 — delivered after payment
— is the open frontier; nobody measures it today, including us. x402station marked UNDETERMINED at
capture (0 services / 0 transactions / 0.0% reliability, so the methodology narrative could not be inspected).
A 402 is a payment challenge, never a buyer. Every surface named is a complementary layer at a
different question, never anyone's error. CC BY 4.0.
Companion reads: probe vs paid (why layers 2 and 3 are strictly nested) · the vetting checklist (the buyer-side five checks) · paid-but-unrated join (the layer-4 dataset).
First buyer-intent piece. You are about to pay an unknown x402 endpoint — here is the checklist we
run against public data before we send money. Five checks, each backed by one of our live numbers.
(1) Has anyone been charged at this endpoint at all? — cohort ladder at ≥1, ≥5, ≥10, ≥100
unique payers in 30 days, rendered live from the same CDP Bazaar snapshot the
market map uses. (2) Diffuse or concentrated?
— exclusive payer-count bucket distribution, so a single-payer endpoint is not confused with a
diffuse one. (3) Does the pay_to walk back to a registry-owned identity?
— the ERC-8004 × x402 ownership join, endpoint AND publisher level in the SAME paragraph, with
the per-publisher dedup cap of 50 disclosed on-page. (4) Is the publisher one endpoint or fifty? (5) Does
any of it carry commerce-backed reputation? Endpoints with a missing or non-numeric
l30DaysUniquePayers field are excluded from every threshold and bucket
— the exclusion note is on-page. Snapshot timestamp is on-page and every 30-day cohort figure is a
ROLLING window; do not read them as permanent numbers. A 402 is a payment challenge, never a
buyer. Probe-liveness measurement is a genuinely useful, complementary layer — the sibling
/blog/probe-vs-paid works through the layer
contrast in depth. CC BY 4.0.
Companion reads: probe vs paid (why the layers are strictly nested) · paid-but-unrated join (the underlying dataset the vetting checklist reads from) · market map (the live population).
There are three different questions you can ask about an agent endpoint, and the ecosystem currently answers them at three different depths — is it advertised, does it answer a probe, was it ever actually PAID. Layer 2 is genuinely useful liveness work — Agents Trust measures it at 138k-URL scale on their landing page (advertised → ever answered a probe → callable now, funnel quoted verbatim with the attribution their page allows: fetched 2026-09-03; no snapshot date given). Layer 3 is a strictly narrower question, and it is the layer we measure: as of the 2026-09-01 CDP Bazaar snapshot, 14,338 earning Base x402 endpoints → 1,291 ERC-8004-owned → 9 rated → 0 commerce-backed, with the publisher-level cut co-located in the SAME paragraph (26 distinct pay_tos, top one 67.2%, 4 rated, 0 commerce-backed; per-publisher dedup capped at 50, disclosed on-page). A 402 is a payment challenge, never a buyer — that is the hinge. Nobody else joins the paid-endpoint cohort to on-chain ERC-8004 identity. CC BY 4.0.
Companion reads: paid-but-unrated join (the layer-3 dataset) · Report 11 (the front-door population cut) · Report 10 (why one-payer endpoints don't count as demand). All CC BY 4.0.
Front door to the x402 content cluster. Every public number circulating for the x402 ecosystem is an
estimate, a projection, or a press figure — this piece is the reference cut we measure
daily from the same live CDP Bazaar snapshot the
market map, the
price census,
Report 07 and
Report 10 render from. CATALOG ≠ PRICED ≠
EARNING kept strictly separate: total indexed endpoints, priced subset with a median call cost that
refutes the circulating "~$0.30 average", earning subset by CDP’s own
l30DaysUniquePayers ≥ 1 field. Payer- and call-weighted concentration
rendered side by side. The ERC-8004 ownership join surfaces at BOTH endpoint AND publisher level in one
section (dated 2026-09-01) with the top-publisher share, distinct pay_tos carrying any feedback, and the
commerce-backed count all co-located; the 0 commerce-backed figure ships with its
Virtuals ACP boundary-condition explanation so it does not read as a data bug. Names what the
population does not carry — reputation, commerce-backing, walkable endpoint identity.
Standing discipline: l30DaysTotalCalls is a request count, not settled
USDC value; a 402 is a payment challenge, not a buyer; the CDP Bazaar catalog is one
Base-weighted x402 discovery source. CC BY 4.0.
Primary datasets (all refresh daily, all CC BY 4.0): market map · price census · paid-but-unrated join · daily catalog time series.
Editorial companion to the call-weighted block that landed on
/x402-bazaar-market-map in task 0423.
Report 07 counted ENDPOINTS.
Report 10 counts CALLS. Same live CDP Bazaar snapshot, opposite picture. The
≥5-unique-payer cohort is a small share of indexed endpoints (with both a payer and a call
count) but carries the overwhelming majority of indexed 30-day requests; the exactly-1-payer
cohort inverts the picture and the MEDIAN 30-day call count for that cohort is 1
— the typical x402 endpoint has been touched about once in thirty days. Every figure
computed at build time from the same snapshot the live market-map concentration block and
the JSON export render from, via the same helper — the page and the live surfaces
cannot drift. Standing discipline: l30DaysTotalCalls is a
request count, not settled USDC value; a 1-payer endpoint is indistinguishable from
self-dealing from on-chain data alone; the CDP Bazaar catalog is one Base-weighted x402
discovery source. CC BY 4.0.
Live surfaces on the same snapshot: /x402-bazaar-market-map — JSON and CSV exports (CC BY 4.0).
Editorial interpretation layer for the 7th citable dataset at /is-the-x402-market-growing. Over 19 UTC days (2026-08-15 to 2026-09-02), the public Coinbase CDP x402 Bazaar catalog's item count moved by a value indistinguishable from zero while distinct hosts grew about 9% and the trough-to-peak intra-window swing was large. That shape is churn, not accumulation — endpoints are being replaced, not added on net. Standing discipline: CATALOG ≠ PRICED ≠ EARNING; do not infer demand from catalog size. One source (the public Coinbase CDP Bazaar), not an ecosystem census. Every figure computed at build time from the daily catalog series cache; CC BY 4.0.
Primary dataset (refreshes daily): Is the x402 market growing? — JSON and CSV exports (CC BY 4.0).
Google announced AP2 on 2025-09-16 as an open protocol for agent-led payments with 60+ launch partners including Coinbase, Mastercard, PayPal, Adyen, and Mysten Labs — the second commoditising payments rail arriving under the same missing trust layer this year, three weeks after Cloudflare Wallets. AP2 authenticates the transaction via cryptographically-signed Mandates backed by Verifiable Credentials, with explicit x402 sample paths in the Apache-2.0 reference implementation. It does not score the counterparty agent. Argued from our own measured data: the 6.5% ERC-8004 → agent-card signpost gap (1,604 of 24,835 indexed registrations), the 97.9% top-10 concentration on the x402 demand side, and Imperial College's arXiv:2606.26028 measurement that the ReputationRegistry cannot function as a trust signal. Fair-comparison table of what AP2's Mandates do attest and what a counterparty-quality layer on top of them still has to prove.
Cloudflare's Wallets announcement (2026-08-04) gives AI agents a
cloudflare.pay identity, a spending cap, and x402 stablecoin micropayments — the
buyer-side half of the internet-native payments stack the Linux Foundation put under a formal x402 Foundation on
2026-07-14. Rails are commoditising. The trust filter is not. Mainstream financial press is now reporting x402
settlement volume down ~93% YTD while payment count keeps climbing — a paradox that only resolves one way: the
scarce good is knowing which counterparty is real. Only 1,604 of 24,835 indexed ERC-8004 registrations (6.5%)
publish a discoverable agent-card; 97.9% of demand-side x402 volume across 423 measured services concentrates in
the top ten. Short fair read of what cloudflare.pay and ERC-8004 each actually
attest, and what a counterparty-quality signal on top of them has to prove.
Measured 2026-08-16: of 24,835 indexed Base + Ethereum ERC-8004 registrations, only 1,604 (6.5%) publish a
discoverable /.well-known/agent-card.json — the standard A2A capability manifest
an autonomous discovery client needs to walk from the on-chain identity to a payable endpoint. Hop-by-hop audit over
seven live agents shows H1 (registry read), H2 (URI resolve) and H5 (V2 402 paywall shape) all 7/7 PASS; H3 is the
single binding hop. Named examples of both the failure (ClawFresh 20992 on Base, Name Whisper 29402 on Ethereum) and
the fix (Tavily 35179, OnchainExpat 23922, Lunairefine 31786, Buyer 32965); a copy-pasteable four-field ERC-8004 v1
reference implementation. We were in the 93.5% at time of measurement — the same-day fix that moved our
/agent.json into strict conformance is documented inline.
OpenSea shipped ERC-8257 (Agent Tool Registry) as a Draft EIP on 2026-04-17 — a permissionless on-chain registry where AI-agent tools publish pricing and agents auto-discover and pay. Live on Ethereum mainnet and Base (roughly 76 tools indexed as of mid-2026, per QuickNode's secondary-source explainer). The three neighbouring layers ERC-8257 assumes — ERC-8004 identity, x402 payments, and MCP capability discovery — are the three we already ship, so registering agent 19353 is bookkeeping. Honest field comparison against agents-trust.ai (Agentscan), which is running close positioning; our differentiator is commerce-backed reputation. On-chain registration queued for Jason.
A June 2026 protocol-governance survey (arXiv:2606.31498, Kang & Diponegoro, DoiT International) scopes reputation out of its six-dimension governance taxonomy on the grounds it is already addressed by ERC-8004. An independent empirical study (arXiv:2606.26028, Imperial College London) measures that same registry and finds it cannot function as a trust signal. Not strictly contradictory — one is a taxonomy scoping decision, the other a measurement of deployed state — but read side by side the honest framing is "the assumption is reasonable on paper and false on chain." Includes the AWS Bedrock AgentCore observation (GA in 15 regions, no trust scoring or reputation) as independent confirmation from the largest cloud vendor that the trust layer is unbuilt.
Two whole-catalog measurements of the x402 agent-payments economy — ours from the supply side (every paid endpoint
in Coinbase's public CDP Bazaar) and x402-list.com's from the demand side (a public /api/v1/stats
endpoint over 423 measured services) — converge on the same shape. A large listing surface, a tiny paying cohort.
In a market this concentrated, a directory of agents is close to worthless and a trust filter is the product.
Third-party figures are frozen citations as measured 2026-08-15; both predicate definitions stated explicitly side by side.
Independent academic work (arXiv:2606.26028, Imperial College London, CC BY 4.0) finds that most ERC-8004 registrations are placeholders and the ReputationRegistry cannot function as a trust signal. Their data froze on 2026-05-13. This post reconciles their frozen May baseline against our live August index, with both predicate definitions stated explicitly side by side. Direction and Base-vs-Ethereum ordering match; BNB remains the weakest of the three chains under both teams' predicates; commerce-backed reputation is the architectural fix for the paper's central complaint about the registry.
Dated whole-catalog measurement of Coinbase's public x402 Bazaar. Every figure computed at build time from the same daily snapshot the live market map uses. The strong-filter cohort (≥5 unique payers in 30 days) is a small minority of the full indexed catalog — listing is not demand. Includes methodology, an honest-correction note, and a citation block under CC BY 4.0.
Primary dataset (refreshes daily): x402 Bazaar Market Map — JSON and CSV exports (CC BY 4.0).
Original data. We downloaded and scanned the entire Coinbase x402 Bazaar catalog to map the agent-intelligence market: who sells ERC-8004 / agent reputation, at what price, and how little anyone is buying yet. At least five direct competitors already ship reputation endpoints; nobody has broken out. First-mover positioning still available. Snapshot as of 13 August 2026.
Live companion (refreshes daily): x402 Bazaar Market Map — same catalog, same normalization, add / remove / price-change delta panel, plus JSON and CSV exports (CC BY 4.0).
Independent research (arXiv 2606.12128) and core Ethereum infra (The Graph, x402, ERC-8004) are converging on the same question: how many registered ERC-8004 agents are actually operational? Our commerce-backed cohort and per-chain live-rate audit are the on-chain dataset that answers it. Complementary empirical evidence alongside the paper's evidence-layer framework and The Graph's cross-chain reputation lookup.
Biggest index is a vanity metric. Live plus commerce-backed is what matters. The chain-quality numbers, the placeholder tail, and the zero commerce-backed count that made cutting BNB the obvious call — and the all-Ethereum posture (Base + Ethereum today, Arbitrum next) going forward.
Our own index of roughly 95,000 ERC-8004 agents on Base, Ethereum, and BNB Chain. Only a few thousand actually answer a request. Per-chain registered-vs-live figures pulled live from our public stats endpoint, with independent research (arXiv:2606.26028) corroborating the shape. Why the registry counts are the wrong number to shop by, and what a real trust filter looks like.
A worked example on agent base/19353: what the free per-agent profile
answers, what the free 200-node graph teaser shows, and what the paid
/v1/intel/peers/{agent_id} and
/v1/intel/graph endpoints add — the ten closest peer agents by capability
overlap, and the full inter-agent payment network in one call.
The six on-chain and off-chain signals that separate a real, usable ERC-8004 AI agent from a paper registration — identity, ReputationRegistry feedback (Sybil-adjusted), x402 and settlement history, cross-chain presence, and live reachability. Vendor-neutral checklist you can apply on any chain.
The small cohort of ERC-8004 agents whose ReputationRegistry rows tie back
to real, paid on-chain jobs (ERC-8183 / Virtuals ACP) — set alongside the total registered set and the
raw feedback-event count so the shape of the gap is legible. The one trust signal a bulk-mint farm cannot
cheaply produce.
Five weeks on: the index is bigger, the BNB coverage moat has widened, and almost none of Report 01's headline agents are still Transact-ready. What the current readiness distribution tells us about which agents actually work.
The inaugural census of the ERC-8004 agent economy — the MCP and OpenAPI capability layers, owner identity, the factory-cluster long tail, and what reachability tells us about which agents actually matter.