Reference
What the rest of the registry is
The catalogue verifies the four categories in depth. This page is about everything else — the tens of thousands of identities that declare a protocol and never appear in a marketplace. It is a sample, and it says so on every number it draws from one.
Measured 5h ago
BSC Mainnet
- 380,175
- registered
- exact
- 48,264
- declare an A2A endpoint
- exact
- 100
- sampled
- across 20% of the population
- 5/10
- strata answered
- the index timed out on the rest
Finding
98 of 100 sampled registrations (98%) declare an endpoint URL with an unsubstituted placeholder still in it — {…} where an identifier should be. Fetched as declared, they return an error. Substituting the agent’s own token id makes 85 of 98 (87%) answer.
That distinction matters. These are not abandoned agents. To be precise about where the defect lives: the on-chain tokenURI resolves fine — it points at a registration document. It is that document whose declared endpoints carry the placeholder. Anyone probing the endpoints as declared — us included, before we checked — counts these agents as dead.
| Token | As registered | With the id substituted |
|---|---|---|
| #413277 | 404 | 200 |
| #413232 | 404 | 200 |
| #413226 | 404 | 200 |
https://platform-backend.prod.termix.live/api/v1/a2a/agents/{agentId}/card
What the hosting platform reports about them
83 · UNBOUND/offline/endpoint:no/skills:0
2 · OFFLINE/offline/endpoint:no/skills:0
Read from the platform’s own response, not inferred from a failed probe. It is a better source about its own agents than our guess would be.
Where the sampled identities point — 2 distinct hosts in 100 identities
Every identity the sample reached resolves to the same backend. That is a statement about the sampled range, not proof about the part the index would not serve.
BSC Testnet
- 2,743
- registered
- exact
- 1,053
- declare an A2A endpoint
- exact
- 200
- sampled
- across 97% of the population
Finding
48 of 200 sampled registrations (24%) declare an endpoint URL with an unsubstituted placeholder still in it — {…} where an identifier should be. Fetched as declared, they return an error. Substituting the agent’s own token id makes 48 of 48 (100%) answer.
That distinction matters. These are not abandoned agents. To be precise about where the defect lives: the on-chain tokenURI resolves fine — it points at a registration document. It is that document whose declared endpoints carry the placeholder. Anyone probing the endpoints as declared — us included, before we checked — counts these agents as dead.
| Token | As registered | With the id substituted |
|---|---|---|
| #2580 | 404 | 200 |
| #2579 | 404 | 200 |
| #2576 | 404 | 200 |
https://platform-backend.dev.termix.live/api/v1/a2a/agents/{agentId}/card
What the hosting platform reports about them
45 · UNBOUND/offline/endpoint:no/skills:0
3 · ONLINE/online/endpoint:no/skills:0
Read from the platform’s own response, not inferred from a failed probe. It is a better source about its own agents than our guess would be.
Where the sampled identities point — 50 distinct hosts in 200 identities
How this was measured, and its limits
Population figures are exact: the registry index returns them as a count. Everything else is a stratified sample, drawn at even intervals across the population rather than from the first page, because the first page and the last turn out to hold different operators.
The index does not paginate reliably to the end. Measured: a request at offset 20,000 takes 19 seconds, and past roughly 22,000 it times out. So the deepest part of the mainnet population is not sampled at all, and the coverage figure above says exactly how much is missing. We would rather publish that than a round number that hides it.
A shared backend is not proof of cloning. A platform that hosts thousands of users’ agents on one domain is doing something ordinary; the number below counts identities per host, not duplicates.
See also how verification works and scope and risk.