Update Nov 2026: we now classify agents with a four-step ladder (Reachable, Transact-ready, Payment verified, Has paying customers). The “Promising”, “Not ready” and “Unrated” labels used here were retired. How the steps work →
Note, 4 November 2026: since 4 November 2026 we label agents on four steps: Reachable, Transact-ready, Payment verified and Has paying customers. This post describes the earlier labels as they stood on its date. See how the steps work.
Part 5
made the case that registered, live, and has reviews from paid jobs
are three different bars, and that shopping by the first one is what
the ecosystem keeps getting wrong. Two months on, the picture has
gained a fourth bar — publishes a discoverable agent card
— and the honest answer for any specific ERC-8004 agent
depends on which of the four bars a buyer needs it to clear.
This post sets the four groups side by side. Where a number is a
sample, or where one publisher's batch could skew a count, we say so
next to the number. It also has a separately reported cut of the
Solana 8004-solana registration
program — a different program that shares a name fragment
and nothing else.
Ethereum registry: raw mints, not agents.
The Ethereum-mainnet ERC-8004 IdentityRegistry
(0x8004A169…539a432) has now
emitted 56,200
raw Transfer(from=0x0) mints as
of the last full block-range census
(2026-11-06). That
number is a count of raw registrations, not agents; it
includes empty rows, factory-cluster tails, single-purpose test mints,
and everything else on-chain minting an agentId costs a few cents to
do. It is a count of raw registrations, not of agents.
Of that raw-mint total,
30,502
(60.9%)
have a populated tokenURI; far fewer
resolve to a usable agent card. A populated URI is a claim that the
minter intended to point at something — it is not
evidence anything is there. The next section reports the fresh
resolvability sample that puts a bounded number on the second half
of that sentence.
The publisher tail matters for reading the mint number honestly. Two
HTTP hosts alone move most of the volume — the top host on
Ethereum by mint count is agents.exquisite.land
at 10,000 mints (roughly a fifth of the whole registry on its own),
and the top twenty hosts together account for roughly a third. Our
per-agent records therefore limit how many agents from one publisher
can appear in our views
(50
per publisher, across
303
publishers; the largest publisher minted
10,000
raw registrations), so one batch mint cannot dominate.
The long tail is part of the story, and we show it rather than hide
it.
The full breakdown, with per-publisher counts, is on our page How many ERC-8004 agents are there?
The resolvability sample: populated URI ≠ working card.
The middle layer of the ladder — between “the on-chain URI is non-empty” and “the agent card is discoverable at the endpoint” — now has a bounded measurement of its own. On 2 September 2026 we drew a stratified random sample of 600 agents per chain from the eligible-nonempty-URI populations (Base: 22,807 eligible; Ethereum: 2,827 eligible) and probed each URI end-to-end, classifying the response into one of nine outcome buckets. The sample is small and stated as such; treat it as a rate estimate over the population, never as a headline census.
| Chain (n = 600 sampled) | Any JSON at URI | Usable agent card |
|---|---|---|
| Base — sample of Base non-empty URIs | 68.17% (95% CI 64.33–71.77) | 56.33% (95% CI 52.34–60.25) |
| Ethereum — sample of Ethereum non-empty URIs | 74.83% (95% CI 71.21–78.14) | 71.33% (95% CI 67.59–74.81) |
| Pooled (n = 1,200) | 71.50% (95% CI 68.88–73.98) | 63.83% (95% CI 61.07–66.50) |
Sample, not census. Full outcome-bucket breakdown, near-miss shape counts, and host concentration by bucket are in the machine-readable dataset on the signpost-gap datasets endpoint and the sibling agent-card signpost-gap post.
Two takeaways. First, on Base roughly 44% of the populated-URI slice does not resolve to a usable agent card in a single probe — the URI was published on-chain but the thing at the other end is missing, wrong, or ghost-DNS. Second, the same slice on Ethereum resolves at a materially higher rate than Base does (71.3% vs 56.3% usable-card). The direction matches the per-chain live-rate shape we have reported since Part 5: Ethereum is a smaller, more curated cohort; Base is larger and more tail-heavy.
The signpost gap: fewer than 6% of registrations carry a discoverable card.
The next bar up is the discoverable
/.well-known/agent-card.json the A2A
protocol assumes on any registered agent. The
signpost-gap dataset
holds a daily snapshot of that predicate for every indexed
(chain, agentId) row: an agent-card is present when the last probe
of the advertised endpoint's
/.well-known/agent-card.json
returned an A2A-conformant JSON body. As of the latest snapshot,
1,524 of 25,695 indexed registrations
(5.93%) publish a discoverable card at all, and
1,231 of 6,146 endpoint-live registrations
(20.03%) do so — the intersection is the
addressable A2A-handshake cohort. That is down
from the 6.5% (1,604 of 24,835) in our
2026-08-16 signpost-gap report.
The number of indexed registrations grew by about 3.5% while the
number with a card fell from 1,604 to 1,524.
Put the other way, about 94% of registrations do not publish the agent card that the A2A (agent-to-agent) protocol expects. An automated client that starts from an ERC-8004 registration will, for most agents, find nothing to read at the next step.
What the index holds today.
Our enriched per-agent records, the rows the paid
/v1/intel/agent API serves, number
26,978.
The distinct agents we profile number
26,978
(23,705
on Base,
3,273
on Ethereum). This is a count of registrations, not of working
agents. The count that matters for a buyer comes next.
Live — endpoint 2xx and a machine-readable capability
surface declared (MCP, OpenAPI, or agent-card) — is
2,483
agents total. Per-chain live rate is
6%
on Base and
33%
on Ethereum, keeping the ordering Part 5 predicted: the smaller
Ethereum cohort is cleaner per capita, Base carries the volume and
the tail. Machine-readable capability surfaces indexed:
9,149
MCP tool descriptions from
712
agents,
17,842
OpenAPI methods from
1,010
agents, and
1,878
agents publishing /.well-known.
The readiness distribution — the four-bucket buyer-POV score we introduced in Part 2 — now covers 18,264 stamped agents. Current split:
- Transact-ready: 62 — live, with a readable x402 price and payTo address, a score of 75 or more, and the anti-spam checks passed. For scale, 291 agents advertise x402 at all.
- Promising: 5,939 — real signals present but at least one binding gap (usually no payment history yet, or an unresolved reputation flag).
- Not ready: 9,020 — unreachable endpoint, expired TLS, or clearly stale.
- Unrated: 5 — registered but not yet fully enriched by the pipeline.
The Transact-ready set remains a small share of the total, even after two months of continuous re-probing. That reflects the ecosystem, not the measurement. The chain-liveness dataset holds the daily per-chain cut of the same signal.
Identity: still the scarcest signal.
Of the 18,586 distinct owner wallets in the indexed store, only 940 resolve to a human-readable identity via ENS, Basenames, or Farcaster. The named ratio has moved little since Part 2 — naming is not scaling with registration. We continue to classify what we can: 3,886 owner contracts are verified on their block explorer, and 593 owners are plain externally-owned accounts. Owner identity is the single largest paid-intel upgrade over the free registry read.
Reputation and the cohort of agents with reviews from paid jobs.
The ReputationRegistry
(0x8004BAa1…19dE9b63) feedback
indexer has now ingested
710,975
events covering
155,829 ERC-8004 agent IDs on the full registry (including registrations outside our index). The overwhelming share is on Base — that is where the
registry was upgraded first and where the deep backfill ran; the
Ethereum tail is a live-tail workload still queued behind an Alchemy
quota budget, so read the reputation numbers as Base-heavy by
construction, not as a claim about Ethereum activity.
The commerce-backed subset — ReputationRegistry rows our commerce-backed definition ties to a paid job outcome (ERC-8183
AgentCommerce hooks or the Virtuals
ACP hosted equivalent) — remains a small set:
51
unique agents,
1,003
job-outcome events, from
117
distinct evaluator clients. The count of commerce-backed agents is small. The comparison that
matters is with the number of agents that have any feedback at all.
Any bulk-mint farm can pay a
cent to mint a registration and a cent to post a five-star review
of itself. It cannot produce a commerce-backed row without first
actually paying an agent through a hook that settles on-chain, at
which point it is a paying customer and the “farm”
framing collapses. The gap between
155,829 agent IDs with any feedback (full registry) and
51 agents with
commerce-backed feedback is the concentration pathology
arXiv:2606.26028 (Imperial College) measures on the same registry,
reported from our indexer with the filter applied. The
commerce-backed-agents
dataset holds the full cohort with the T1/T3 hook classification
per row.
Solana: a different registry, reported separately.
Solana runs its own agent-registration program —
QuantuLabs's
8004-solana
(8oo4dC4JvBLwy5tGgiH3WwK4B9PWxL9Z4XjA2jzkQMbQ).
It shares a name fragment with ERC-8004 and nothing else: different
chain, different program, different account model, different
semantics. It is not an ERC-8004 registry, so its figures should be
reported separately from Ethereum registry figures. Our 2 September 2026 census of that
program's accounts, run against public Solana mainnet RPC and
published verbatim as a snapshot to the
Solana registry census page,
found 1,477 raw registration accounts from
185 distinct creators, with the top-1 creator
holding 22.07% of accounts, the top-10 holding
78.67%, and the top-20 holding 85.65%.
99.4% of accounts have a populated
agent_uri field, but the census
does not fetch any URI and therefore makes no resolvability
claim about the Solana rows — a populated URI is a claim of
intent, not a claim of a working card.
Report those figures alongside the Ethereum registry rows if you want, but not as Ethereum registry rows. Concentration this high (top-20 = 85.65%) is the shape of a registry a small number of publishers batch-minted into; anyone building trust on it should read the creator histogram, not the account total.
What did not change, and why that is a finding.
Three things stayed flat against Part 5. The named-identity
ratio (owner wallets with an ENS / Basename / Farcaster tag) barely
moved. The Transact-ready share stayed a small fraction of the
indexed total. The commerce-backed cohort grew by a couple of
agents at most, not a couple of orders of magnitude. In an
ecosystem where the raw registration count is climbing steadily,
those three flat lines are the finding: registration cost fell to
a rounding error long enough ago that the tail is now the median,
and every filter downstream of raw Transfer(from=0x0)
counts is doing more work than it was two months ago.
Across the series the pattern is consistent: the useful number is a small, filtered, continuously verified group, not the large registration total. Headline counts are live; every sample or dated snapshot is labelled in the section that uses it. Machine-readable versions of each dataset are linked below.
Part 5 asked what “operational readiness” looks like
for one ERC-8004 agent. Part 12 answers the same question two
months later at population scale: registered is cheap, signposted
is scarcer, resolvable is scarcer still, and paid is the honest
bar — and the four cohorts are not the same set of agents even when they overlap. The full per-agent evidence for any specific
(chain, agentId) is available live at
/v1/intel/agent.