All posts
AgentsERC-8004

How BAS checks 8004 agent

A registration says an agent exists. Doctor asks whether anything answers, what it answers with, and how many distinct services stand behind the count.

BAS Team

A small dog knight with a shield stands in a flower meadow beside a pale dragon, under the words How BAS check 8004 agent

Headline: 941,487 ERC-8004 agents, 977 that show their own card: how BAS Doctor checks an agent

Subtitle: A registration says an agent exists. Doctor asks whether anything answers, what it answers with, and how many distinct services stand behind the count.

TL;DR

  • BAS Doctor checks the ERC-8004 agents BAS indexes and grades each one by what answered, never by what its card declares.
  • In the run of November 12, 2026, Doctor considered 941,487 registrations. 618,724 declared an endpoint it contacted, and 517,314 answered.
  • Removing registrations that only repeat another’s response leaves 359,669. Of those, 977 serve their own agent card: about 1 in 960 registrations.
  • Registrations are not suppliers. The 349,022 that answer on an agent protocol sit on 349 hosts; in a random sample of 1,200 of them, all 1,200 were on one host.
  • Every check is read-only, one request per host at a time. Anything Doctor could not check is reported as not checked, never as a pass.

Registering is free; answering is not

ERC-8004 lets anyone register an agent on chain and point it at a registration card. Listing endpoints on that card is optional: the standard says agents MAY advertise them. A registry count therefore measures registrations, not working agents.

Doctor’s rule fits in one sentence: a level describes what answered, never what was declared. An agent that lists six endpoints and answers on none sits on the same rung as one that never tried.

The rule exists because declaring costs nothing. In the same run, 47,730 registrations still had a template placeholder such as {agentId} in an endpoint, and 84,585 had no card address on chain at all.

The ladder: L0 to L3

More than half of all registrations answer something, but fewer than 2 in 1,000 answer with a real agent card.

LevelWhat it meansRegistrationsShare
L0Registered; declares no endpoint322,72234.3%
L1Declares an endpoint; nothing answered, or it was not contacted101,45110.8%
L2Something answered515,55454.8%
L3An endpoint served its protocol's own machine-readable descriptor: an A2A agent card or an MCP tool list1,7600.19%

L3 is a real bar, not “returned JSON”. Doctor checks that the body actually is a card. In an early survey, counting any JSON reply as machine-readable gave a 9.7% pass rate; checking the body gave 0%.

L1 has two causes, and Doctor keeps them apart: silence, and an endpoint nobody knocked on. In this run, 41 registrations declared an endpoint that was never contacted. They are reported as unasked, not as silent.

From 941,487 registrations to 977 cards

Of 941,487 registered agents, 977 serve their own agent card

StepMeaningRegistrationsShare of the step before
Consideredevery registration with a Doctor report941,487starting point
Contacteddeclared an endpoint Doctor dialled618,72465.7%
Answeredat least one endpoint responded517,31483.6%
Listedafter folding repeated responses359,66969.5%
Agent protocolanswers on A2A or MCP349,02297.0%
Own cardserves its own agent card (L3)9770.28%

BAS 8004 Doctor registry roll-up · public shortlist API · Nov 12, 2026, 07:20 UTC

The steep drops are at the two ends. A third of all registrations declare nothing to contact. At the other end, 349,022 answer on an agent protocol, but only 977 serve a card that describes the agent.

Each step is measured against the step before it. A rate that divides what answered by everything registered would understate the answer rate and overstate how much of the registry was actually asked.

One response, one service

Doctor hashes every response body it receives. A registration is listed only if it is the first to serve at least one body; one that only repeats what another already served is folded into it. In this run that removed 157,645 of the 517,314 that answered.

The same idea keeps the checks polite. When 20 different URLs of one shape on one host return one identical body (think /agent/detail?id=N for any N), Doctor stops probing that shape: the per-agent part of the URL selects nothing. It stops the URL shape, not the host, so a platform that serves real per-agent cards at another path is still checked there.

Body hashes catch copies, not templates. A platform that writes each agent’s id into an otherwise identical reply produces a distinct body per agent. So Doctor also counts hosts, and the gap is large:

TierRegistrationsDistinct responsesHosts
Good: serves its own card, or has been paid977883178
Usable: answers on an agent protocol (A2A or MCP)349,022348,910349
Reachable: anything answered359,669353,834818

In a random sample of 1,200 Usable registrations, all 1,200 were on one host. Read the Usable count as mostly one platform, not as 349,022 suppliers.

What Doctor never does

Every endpoint Doctor contacts was written by a stranger, so the checks are built to be safe to run against anything:

  • Read-only requests only. Doctor sends GET requests. Any request body it can send is a constant in its own source, for a call the protocol defines as read-only, so a card cannot make it send something else. It never calls an agent’s tools.
  • One request per host at a time. One operator can hold hundreds of thousands of registrations behind one hostname. Probing them all at once would look like an attack.
  • It checks the address it actually connects to. An early run found registrations pointing at localhost. Requests to private or local addresses are refused and reported as blocked.
  • It never keeps a credential. When a card contains something shaped like an API key or a private key, the report names the pattern and its position, not the value.

A gap never reads as a pass

A single-agent report always has the same six sections. A section Doctor could not produce still appears, with the reason:

  1. Conformance: does the registration follow ERC-8004? Each verdict cites the sentence of the standard behind it, and MUST, SHOULD and MAY are kept apart.
  2. Hygiene: is it broken as written?
  3. Liveness: does anything answer? The only section that sends requests.
  4. Claimed vs tested: does the card contradict itself, or what was observed?
  5. Security: what holds without invoking the agent.
  6. Settlement: has anybody paid it through ERC-8183?

Each section carries one of five status words: complete, partial, not checked (we did not look), not applicable, or absent (we looked; the optional thing is not there). They are not synonyms. “Not checked” and “nothing found” lead a reader to different decisions.

Two rules follow from that:

  • No claim is ever reported as verified. The strongest verdict is “consistent”: nothing we saw contradicts it. A card saying x402Support: true stays untested, because confirming it means sending a paid request.
  • The security section is never complete. It lists what it did not test: prompt injection, authentication, rate limits, and what the agent does with the data it is given.

Hygiene flags show how many registrations are broken before any request is sent:

FlagWhat it catchesRegistrations
Empty card addressNo agent URI on chain, so there is no card to read84,585
Missing registrationsThe card never says which registry it belongs to69,451
Template placeholderAn endpoint still contains {agentId}47,730
Unsupported card addressA URI scheme Doctor will not follow5,539
Address instead of URLA chain address such as eip155:56:0x… in an endpoint field2,770
Standard as endpointA link to the EIP pasted into a service1,519
Standard as cardThe agent URI points at the EIP, not at a card1,442
Card over plain httpAnyone on the path could have rewritten it793
Third-party domainAn endpoint on a domain the registrant does not control640

What Doctor cannot tell you yet

Four limits, stated so the numbers above are read correctly:

  • MCP agents are not contacted. Checking an MCP endpoint takes a JSON-RPC POST, and the scheduled runs do not send one. MCP endpoints come back “not checked”, so the L3 count above is A2A cards only.
  • The registry-wide view does not carry payments yet. Its “paid on-chain” line reads “no settlement evidence was supplied”, not zero. The single-agent report does read ERC-8183 history, with its own rules: no success rate over a single job, and jobs that all share one counterparty are labelled as such.
  • Trust models are taken as declared. Most cards that declare one declare reputation, but ERC-8004 feedback costs nothing to write, so Doctor does not count it as evidence. Staked validation and TEE attestation could be checked on chain; Doctor does not check them yet.
  • Checks are spread over days. Agents that serve a card are re-checked daily, agents that answer every 3 days, silent ones on a back-off from 1 to 14 days, hosts that have gone every 30 days. At this snapshot, 95% of contacted agents had been checked in the last 10 days; the oldest check was 72 days old.

Check an agent yourself

Everything above is live on agent.bas.io:

  • Doctor page. agent.bas.io/doctor , titled “Agent Doctor”, shows the funnel, the registry’s levels and hygiene flags, and the Good, Usable and Reachable lists with registrations and hosts side by side.
  • Any agent. In “Check any agent — uid or agent URL”, a uid opens that agent’s Diagnostics tab, and a card URL gets a report on the spot, even for an agent outside our index; checks that need to know who owns the card, such as payment history, then come back “not applicable”.
  • Agent page. Every agent page has a Diagnostics tab with the six sections.
  • API. /agent/api/doctor/report?uid=N returns one report; /agent/api/doctor/shortlist?tier=good returns a tier together with the funnel.

How we counted

  • Snapshot: Doctor’s registry roll-up of November 12, 2026, 07:20 UTC, read from the public shortlist API.
  • Considered: every registration Doctor holds a report for, across the chains BAS indexes.
  • Contacted: at least one declared endpoint was actually dialled and its outcome recorded.
  • Answered: at least one endpoint responded.
  • Listed: the first registration to serve at least one response body. Later registrations that serve only bodies already seen are folded in.
  • Agent protocol: the answering endpoint is declared as A2A or MCP. Website, GitHub and social links count as answered, never as an agent protocol.
  • Own card (Good): listed, on an agent protocol, and at L3. Being paid through ERC-8183 also qualifies, but this snapshot carried no payment data.
  • Hosts and distinct responses: per tier, from the same API. The 1,200-row sample is 12 random pages of 100 rows from the Usable list.
  • Earlier survey figures (the 9.7% against 0%) come from Doctor’s development runs.