On-chain identity vs off-chain information

When you run into an agent, where do you actually find its identity?

Eight months after mainnet, we wanted a narrower question than “how many agents exist.” Where does a readable agent card live — on the registry, in the URI file, or nowhere?

This is what we used:

• Snapshot date: 24 August 2026.
• Universe: Identity on mainnet only, on eight EVM chains: Ethereum L1, Base, Polygon, BNB Chain, Arbitrum, Celo, Gnosis, X Layer.
• Stock: 441,794 agents.
• “Usable card”: a non-empty name and a description of at least 10 characters. A name alone is a label, not a card.

On-chain information

On-chain field	Agents	% of 441,794
Has a chain profile	441,794	100%
Non-empty name	441,794	100%
Description ≥ 10 characters (usable card)	205,564	46.5%
Name without a usable description	236,230	53.5%
Empty description	234,253	53.0%
Description of 1–9 characters	1,977	0.4%
Non-empty image	172,868	39.1%
Name + description + image	171,269	38.8%
Services, skills, or tags on the chain profile	0	0%

Off-chain information (URI file)

Off-chain URI	Agents	%
Declares an agentURI	391,235	88.6%
Parsed URI profile	371,473	84.1%
Usable URI profile (name + description ≥ 10)	352,017	79.7%
URI profile with an image	306,056	69.3%
URI profile with services	181,132	41.0%
Document stored behind the pointer	387,211	87.6%

Off-chain information (feedback)

Feedback-origin profile	Agents
Any feedback-origin profile	9,463 (2.1%)
Usable name + description	5,456 (1.2%)

Three origins, one gap

Bucket	Agents	%
Usable on-chain card	205,564	46.5%
URI profile (any fill)	371,473	84.1%
Feedback profile (any fill)	9,463	2.1%
Usable chain and URI profile	204,323	46.2%
All three origins	3,983	0.9%
Usable chain only (no URI, no feedback profile)	1,240	0.3%
No usable chain, URI profile only	163,910	37.1%
No usable chain, but URI or feedback profile	169,389	38.3%
No usable chain, no URI profile, no feedback profile	66,841	15.1%

Same standard, different cards

Chain	Agents	% of stock	On-chain usable	URI profile	URI usable	Feedback profile
BNB Chain	296,616	67.1	52.6%	94.7%	93.8%	0.2%
Base	67,701	15.3	42.1%	64.9%	57.6%	8.9%
Ethereum L1	50,397	11.4	21.2%	40.4%	29.8%	0.9%
X Layer	11,170	2.5	0.3%	97.0%	42.3%	0.0%
Celo	9,779	2.2	90.1%	98.0%	96.7%	14.5%
Gnosis	4,115	0.9	6.4%	99.5%	96.9%	24.1%
Arbitrum	1,387	0.3	86.7%	97.4%	91.1%	2.1%
Polygon	629	0.1	13.5%	82.7%	67.7%	14.1%

Analysis

The previous note asked where ERC-8004 identities and reputation live. This one is narrower: when you meet an ERC-8004 agent, where is its identity actually stored?

What the standard designed. Identity is an ERC-721 plus a pointer. From the start it was meant to be dual: some of the identity on the contract, the rest in an external file where the agent has more room for skills, services, and endpoints. Feedback can even attach a third pointer, which sometimes holds the same document and sometimes another.

The numbers match that design: 100% of agents have an on-chain name, and more than 84.1% have a parsed URI profile.

What the data show. As before, you have to go one layer down or you stay in the spec.

1. The registry is a directory, not a catalog. An on-chain name for 441,794 agents sounds like transparency. In detail, more than half of those names have no description long enough to read, and no chain profile in this index carries services, skills, or tags.

   If you only look at the on-chain record, about one in two agents will look like a dummy or get discarded for lack of information. Measuring the quality of the ~50% that do write enough is a later note. The headline already is that metadata quality is an alarm.

   Stay on the NFT and you learn the agent exists. You do not learn what it does.

2. agentURI is a pointer — for better and for worse. An off-chain pointer gives the contract a freedom it does not have. It also puts the burden on whoever hosts it: keep it up, and keep it possible to check that the file is still there and still current.

   That already shows. About 88.6% of agents declare an external pointer. 79.7% have a usable name plus description in that file. Ask for an image as well and it drops to 69.3%. Ask whether the agent declares what it does (services) and only 41.0% do.

   Great power, great responsibility.

3. Chain culture does more work than the spec. Celo writes a usable on-chain card for 90.1% of its agents. X Layer does it for 0.3% and still sets a URI on 97%. Ethereum L1, which opened the year as the only venue, now has the weakest dual coverage: 21.2% usable on-chain and 40.4% with a URI profile. BNB, which holds 67% of the stock, keeps the split the standard imagined: a half-filled NFT and an almost complete URI file.

   In one ERC-8004 ecosystem, each chain’s habits — and often the requirements of its main marketplaces — show up in how local agents register.

   We copy the habits of whoever we follow.

Conclusion

On-chain identity is not mature enough, on its own, to let you know the agent you want to interact with.

That is a discoverability problem. An agent that only reads the registry hits a hard wall, and ends up needing tools that fetch the off-chain agentURI.

It also pushes agent marketplaces to keep their own processes for reading agentURI and filling in services, so agents can actually find each other.

And a third fact: any tool that wants to analyze or verify agents has to read off-chain to complement what each agent claims to have, and keep extra processes to check that the file is still current and reachable.

Fold in the earlier map of ERC-8004 growth and the standard has already moved through stages: first a mint surface, then a reputation surface, and now a documentation surface that lives off-chain. Read only the registry and you undercount what agents say they are. Read “441k agents” as 441k complete cards and you overcount what the chain knows by itself.

What this does not show

• That a parsed profile is live, honest, or unique. Present is not reachable, and it is not a quality score.
• That the 15.1% “with no card” do not exist: all of them have an on-chain name. What is missing is a usable description or a parsed URI/feedback profile.
• Clones, templates, or richness. That is a different reading.
• Validation. The contract is not on mainnet.
• Every ERC-8004 deploy — only the eight mainnets in this index.

How these numbers were cut

Snapshot 24 August 2026, ~04:08 UTC. Identity on mainnet, eight EVM chains: Ethereum L1, Base, Polygon, BNB Chain, Arbitrum, Celo, Gnosis, and X Layer. Testnets are out.

A “usable card” is a non-empty name and a description of at least 10 characters. Origins are counted from profile rows (chain, parsed URI, feedback), not from a merged view. A document stored behind the pointer means bytes were saved; it is not a liveness check.

Related read, not republished here: eight months of ERC-8004 on mainnet.

Data indexed by Global Score Agent. This is a reading of ERC-8004 identity documents, not a product scorecard.