How many AI agents already share a name?
The Agent Name Census reads named agents across the major on-chain registries and measures the two things nobody else can: how often names collide, and how often they land on a live US trademark. Public method, pinned block heights, reproducible. Bounded samples across Ethereum, Base and BNB are published below.
The Census enumerates named agents across Ethereum, Base and BNB ERC-8004 registries, deduplicates by owner, and computes name-collision rates (exact and by sound) and US-trademark knockout rates on the most-duplicated names. It reports collision signals with a stated method — never a legal opinion on any individual agent.
First bounded sample — Base mainnet
These are real, block-stamped measurements from a first bounded window, not projections. Here is the exact scope, so each number means precisely what it says: the ERC-8004 Identity Registry 0x8004A169…a432 on Base mainnet, ERC-721 mint logs over blocks 43,540,847–43,550,846 (a 10,000-block window), read on 2026-07-27. One chain, one deep window. The second sample below widens coverage to Ethereum and BNB over shorter windows; wider windows on all three remain the next edition.
| Measure | What it counts | Result |
|---|---|---|
| Registrations read | ERC-721 agent mints in the window | 136 |
| Named agents | cards resolved to a name, deduped by (name, owner) | 78 |
| Duplicate registrations | same (name, owner) minted more than once | 4 |
| Dead identities | cards that 404'd or failed to fetch — recorded, not dropped | 54 |
| Exact-name collisions | one name used by >1 distinct owner | 0 in window |
| Exact live US-mark hits | resolved names matching a live US-registered wordmark | 11 of 78 |
| Hard trademark bar | names hitting a goods-independent bar (famous-mark dilution) | 1 — “Claude” |
The accounting is closed: 78 named + 4 duplicate + 54 dead = 136 registrations read — every mint in the window is accounted for. Of the 78 resolved names, 11 exactly match a live mark on the US register (for example forge, pixel, aegis, scout) and one — Claude — hits a famous-mark dilution bar. The remaining names return CANNOT_DETERMINE for US confusion, not “clear”: ERC-8004 agent cards carry no goods description, so the confusion analysis can’t complete — reported as unknown, honestly, never as a false all-clear.
Status: first sample published and dated. This page is the methodology of record; each edition is dated, block-stamped and archived, and the numbers above expand as Ethereum, BNB and wider windows are added.
Three chains — Ethereum, Base and BNB
This second sample widens coverage from one chain to all three — real, block-stamped reads on the same registry 0x8004A169…a432, run 2026-07-27, over shorter windows. Exact scope, so each number means what it says: ERC-721 mint logs over
- Ethereum — blocks 25,620,383–25,623,382 (3,000 blocks)
- Base — blocks 49,176,904–49,179,903 (3,000 blocks)
- BNB — blocks 112,418,526–112,420,525 (2,000 blocks; the next 1,000-block chunk was rate-limited by the public RPC and left unread — recorded as a finding, not counted as scanned)
| Measure | What it counts | Result |
|---|---|---|
| Registrations read | ERC-721 agent mints across the three windows | 24 |
| Named agents | cards resolved to a name, deduped by (name, owner) | 19 |
| Distinct names | unique names among those 19 agents | 13 |
| Exact-name collisions | one name minted by >1 distinct owner | 2 names |
| Duplicate registrations | same (name, owner) minted more than once | 1 |
| Dead / malformed identities | cards that failed to fetch or parse — recorded, not dropped | 4 |
| Famous-mark hard bars | names hitting a goods-independent famous-mark bar | 0 in window |
| Exact live US-mark hits | requires the full ~14M-mark register index — deferred pass | not run here |
The accounting is closed over what was read: 19 named + 1 duplicate + 4 unreachable = 24 registrations. By chain: Ethereum 8 read (6 named), Base 3 read (0 named — all three cards set a non-URL identity, each recorded as a finding), BNB 13 read (13 named).
The headline is the collision: two names are already shared across distinct owners in these small windows — Ave.ai Trading Agent minted by six different BNB owners, and Signalbound #1197 by two different Ethereum owners. That is the exact pattern the Census exists to surface, and it did not appear in the single Base window above.
What did not run here, stated plainly: the exact live US-register knockout — the pass that produced the Base edition’s 11-of-78 — needs the full ~14M-mark index and is a deployed-engine step, not part of this in-window sample, so it is reported as not run, never as “clear.” The index-free axes did run: zero names hit a famous-mark dilution bar in this window, and a multilingual offensive-term screen (a pattern signal, never a determination on any agent) flagged a few names for human review. Status: second sample published and dated; coverage widens and the full US-register pass is folded in as later editions.
How the Census is built
- Enumerate — the ERC-8004 Identity Registry singleton (0x8004A169…a432) is pinned from the reference repo (README + addresses.ts) and confirmed against on-chain bytecode on all three mainnets; ERC-721 mint logs are read in fixed block windows at a recorded height.
- Resolve — each agent card fetched; unreachable or nameless cards are recorded as findings, not dropped.
- Deduplicate — by (name, owner) before any collision is counted, so one owner’s reuse isn’t mistaken for a clash.
- Screen — exact and phonetic collision across the corpus; US-trademark knockout on the most-duplicated names only (the two axes that scale).
- Publish — findings as a dated dataset with the block height and method attached.
Scope and ethics. The trademark knockout is US-only, and stated as such. No personal data is republished, and no individual builder is named or shamed — the Census reports rates and patterns, with a method, never a verdict on a person’s agent.