Two things landed in ERC-8004 world at almost the same moment. The first is an arXiv preprint — 2606.12128, From Agent Identity to Agent Economy: Measuring the Operational Readiness of ERC-8004 AI Agents by Rischan Mafrur and Priagung Khusumanegara — that names the exact quantity we've been treating as the load-bearing number since Report 01: not registration count, not attestation count, but operational readiness. The second is The Graph Foundation's write-up of x402 and ERC-8004 as paired primitives, framing indexed cross-chain reputation lookup as the practical shape of an inter-chain agent economy.
Both pieces circle the same problem from opposite directions. Independent researchers are asking "how many of these registered agents are actually doing anything?" and core infra is asking "what do you need to build for the ones that are?" The register-vs-live gap stops being a niche complaint and becomes the shared frame for the next wave of ERC-8004 work.
We publish the on-chain dataset that answers it. This post takes both of those external artifacts at face value, cites what they actually say, and shows where our commerce-backed cohort and per-chain live-rate audit fit alongside them as complementary empirical evidence.
What the arXiv paper actually measures
From Agent Identity to Agent Economy studies the first ten thousand ERC-8004 agent IDs on Ethereum mainnet over a bounded block-range window (roughly 29 January to 9 April 2026). Rather than proposing a single "operational" flag, the authors build a layered evidence framework and score every agent zero-to-five on observable signals: metadata availability, service endpoint declarations, ReputationRegistry feedback records, cross-chain registration, and transfer activity. The methodology combines feature engineering with network analysis over the owner-agent, feedback-client, and wallet-transfer graphs.
The paper's headline finding is that early adoption is "registration-heavy but operationally shallow". The published funnel from the abstract and body:
- Only 67 agents (0.67%) declare service records.
- Only 628 agents (6.28%) have any ReputationRegistry feedback at all.
- Only 19 agents (0.19%) satisfy every layer of the framework — metadata plus services plus feedback plus cross-chain evidence.
- The top ten owner wallets hold 51.4% of the full ten-thousand-agent slice. The single largest feedback client accounts for 65.8% of all feedback records.
The paper is empirical rather than prescriptive. The authors do not pick winners; they describe a distribution. But the distribution they describe is the same one every serious integrator has been running into for months: raw registration numbers are cheap and misleading, and the honest signal lives in a much smaller cohort behind several layers of evidence.
Two things about this we have been publishing continuously from a
different angle. First, "service records" in the paper's sense —
an agent that declares a service — is a subset of the
stricter test we run in production, where an agent must not only
declare an endpoint but also return 2xx on it and expose a
machine-readable surface (MCP, OpenAPI, or
/.well-known/agent-card.json).
Second, the paper's "reputation feedback" count is unfiltered by
Sybil concentration — the 65.8% concentration in a single feedback
client is exactly the pathology our
commerce-backed
predicate was designed to filter around. The paper's raw
feedback figure is the pre-filter number; the commerce-backed
figure we publish is the post-filter number for the same registry.
What The Graph Foundation's post frames
Understanding x402 and ERC-8004 (The Graph Foundation, 5 February 2026) takes the paired standards at their intended use case. ERC-8004 provides verifiable on-chain identity for agents; x402 provides the HTTP-native micropayment settlement layer that lets those agents transact autonomously. The post's operational angle is that indexed cross-chain reputation lookup is what makes the pair usable in practice — "an AI agent operating on Base can instantly verify the reputation of an agent on Arbitrum by simply querying a Subgraph."
The post covers eight chains, with Base and Arbitrum called out as the working examples. The Graph's role in the framing is the data infrastructure that makes cross-chain reputation queries fast enough to matter, and GraphTally is the micropayment layer that makes the indexer economics work at agent scale.
Two takeaways for anyone building on top. One: the ERC-8004 + x402 stack is being positioned as inter-chain by default, not single-chain-first. Two: fast, reliable indexed access to the underlying registry state is the piece the ecosystem is treating as the missing primitive.
Both are directly downstream of the register-vs-live-vs-paid distribution. A cross-chain reputation lookup that returns "yes, this agent has reputation" is worth much less if the agent behind the row cannot answer a request or has never been paid. The indexer is necessary; the trust filter on top of it is what makes the answer useful.
Where our data fits
Our per-chain, live-probed dataset is the direct empirical complement to both pieces. The paper counts registrations, declarations, and feedback rows across one chain and one 10k slice; our index runs continuously across two chains and pairs each registered row with an active liveness probe and a commerce-backed predicate:
| Measure (live, all-Ethereum index) | Value |
|---|---|
| ERC-8004 agents indexed (Base + Ethereum) | 25,618 |
| Live agents (endpoint 2xx AND capability declared) | 3,323 |
| Live rate — Base | 7% |
| Live rate — Ethereum mainnet | 58% |
| ReputationRegistry feedback events indexed | 277,710 |
| Agents that have received any feedback at all | 29,128 |
| Commerce-backed agents (got-paid cohort) | 44 |
| Job-outcome feedback rows (T1 / T3 evidence) | 979 |
Live-source: /v1/public/stats.
Liveness itself has been continuously tracked since
2026-07-17.
The two datasets sit next to each other cleanly. The paper's
"service records" layer is the declared-endpoint predicate; our
live number is that same layer verified by a real
request. The paper's raw feedback count is the ReputationRegistry
state pre-filter; our commerce-backed number is that same
state post-filter, where the filter requires the counterparty to
have actually paid the agent through an ERC-8183
AgentCommerce job outcome or the
Virtuals ACP hosted equivalent. The concentration pathology the
paper documents (one client responsible for 65.8% of all feedback)
is why the raw count is not the number to shop by — a Sybil farm
can produce endorsements at zero cost, but cannot produce a
commerce-backed row without first paying the agent, at which point
it is a paying customer.
Complementary, not competitive
The arXiv authors are doing academic measurement. The Graph Foundation is doing infrastructure positioning. We publish a live, commercially operated dataset that either can build on and cite. The three angles are not the same product; they answer three different questions:
- The paper answers "what does the observable evidence distribution look like on Ethereum mainnet for the first 10k agents?"
- The Graph post answers "what does inter-chain reputation lookup need from an indexer to be practical?"
- Our index answers "for any specific agent right now, on Base or Ethereum mainnet — is it live, has it been paid, and who has it worked with?"
The through-line is operational readiness. Registered is the cheap bar; live is the medium bar; commerce-backed is the honest bar. Any downstream product that has to pick agents to hire, route payments to, or trust with jobs has to get past all three.
Three commerce-backed agents you can inspect right now
Every one of these free per-agent pages carries the readiness
bucket, live-endpoint composition, capability tags, and
commerce-backed status the paper's framework would score on.
Follow the links and see the same signals a paid
/v1/intel/agent response is
built from:
- Lunara (Base) — one of the deepest job-outcome rows in the current commerce-backed cohort.
- Ethy AI (Base) — hosted ACP integration with an actively growing job trail.
- Capminal (Base) — another top-of-cohort commerce-backed profile.
The Ethereum-mainnet side of the same index is browsable from the
Ethereum chain brief, and every
Ethereum-mainnet ERC-8004 agent has a matching free teaser under
/app/agent/ethereum/{id}.
The trust filter, not the registry dump
The line the series has been walking since
Report 01 is the same
one the paper and The Graph post are now walking from their own
angles. Anyone can count
Transfer(from=0x0). Very few can
publish, for any specific agent, a probe-verified liveness answer
and a commerce-backed reputation status. The paper's 0.19%
"full-evidence" figure and our commerce-backed cohort are the same
story told with two different filter stacks — and both converge
on the same conclusion: the useful signal in ERC-8004 is a
small, filtered, observable slice, and
surfacing it is the piece the ecosystem now needs to build around.
The same trust-filter cut applied to the wider x402 paid-endpoint
catalog lives on our
live x402 Bazaar market map,
which buckets Coinbase's public CDP discovery API by unique-payer
thresholds over the last 30 days — the same "is anyone actually
paying this endpoint" question, one layer up.