Who claims to be an agent, who vouched for them, and how much of either turns into anything measurable. A registration is a claim, not activity.Note 23
Three separate records of the same population, and the distance between them is the finding. A registration is something an agent says about itself; a rating is somebody else saying something; a settlement is money moving.
These are not three stages of one funnel. An agent can be rated without being paid on any rail we read, and paid without ever being rated, so they are three populations, not three steps, and a conversion rate between them would be a number we invented. No registered owner yet appears on a settlement rail we read. That is a floor, not a count: the identity registry publishes no operating address other than the NFT owner, so an agent working from an address it never declared is invisible to this join.
Read the rating counts with the rater counts. On at least one network the ratings come from a handful of addresses publishing hundreds each, against five or six per rater elsewhere. A reputation total produced by sixty-five parties is a different object from the same total spread across thousands, and only the second is evidence that anybody independent looked.
| Network | Registered | Rated | Share rated | Ratings per rater |
|---|---|---|---|---|
| BSC | 341,321 | 4,444 | 1.3% | 266from 112 |
| Base | 85,543 | 29,781 | 34.8% | 35from 13,290 |
| Ethereum | 50,746 | 1,676 | 3.3% | 5.1from 664 |
| Abstract | 16,379 | 166 | 1.0% | 15from 20 |
| 13,288 | 1,391 | 10.5% | 5.7from 3,821 | |
| Celo | 5,046 | 2,025 | 40.1% | 5.5from 1,540 |
| Gnosis | 4,127 | 2,081 | 50.4% | 4341from 1 |
| 2,098 | 0 | 0 | — | |
| Avalanche | 1,815 | 1,766 | 97.3% | 1312from 11 |
| Arbitrum | 1,477 | 93 | 6.3% | 29from 10 |
| Polygon | 639 | 157 | 24.6% | 32from 15 |
| Optimism | 542 | 7 | 1.3% | 83from 2 |
| Tempo | 289 | 260 | 90.0% | 4.7from 255 |
| Robinhood Chain | 245 | 2 | 0.8% | 1.0from 2 |
| Mantle | 144 | 18 | 12.5% | 131from 13 |
| Metis | 138 | 0 | 0 | — |
| Linea | 115 | 23 | 20.0% | 31from 1 |
| Scroll | 108 | 0 | 0 | — |
| Taiko | 94 | 0 | 0 | — |
ERC-8004 identity registry · joined against the two rails above
ERC-8004 identity registry
Shares of every registration we read, across 19 chains. The ring folds anything under one per cent into one slice; the exact figures are in the table on this page.
Registrations do not connect to earnings in the markets we read. Read the middle figure as a floor set by the narrower side, not as a rate at which registration converts into work.
A registry counts registrations and a venue counts its own users; joining the two is what says whether a cohort has participants or announcements. The join is ours; the two sides are public.
The two sides are not read over the same networks. Registrations cover four; the markets below cover two rails, one of them on a single network. An agent registered on a network we read no market for cannot appear in the second number however much it earns.
The join is on the address that holds the identity. The registry does publish a separate agent wallet, and we called it rather than assuming: getAgentWallet answered for a sample spread across the whole range of registered ids and returned the owner every time, with no case of the two differing. So for this cohort the holder address is the declared wallet, and the join is not losing agents there.
What it is bounded by is the other side. This reads two rails over one window, and an agent paid anywhere else is invisible to it, On-Chain Agent Intel, reading a wider commerce set, reports dozens of registered agents that have been paid. Both figures can be right, because they are answers to different questions, and only one of them is on this page. Read the number above as these registrations do not connect to earnings in the markets we read, never as a count of what the ecosystem pays.
ERC-8004
524,154 registrations read. Every one is exactly one of these three.
36% of registrations put the card on the chain, 50% give a URI somewhere else that we do not follow, and 14% give neither.
Only the last of those three is a statement about the registry. 71,216 carry neither a card nor a pointer we could parse, and those are unreadable to any client rather than merely to ours. The middle group may well resolve perfectly; we do not fetch them, because a count that depends on a gateway answering is not a count, and that is a limit of this reading.
The two are worth keeping apart even though together they make a more striking figure. Summed, they say most of a machine-readable registry is unreadable; separated, they say a third of it is readable without asking anybody, and about half of it asks you to trust a server.
A card held as an ipfs:// or https:// link is a claim about where a claim lives. Following one would make this count depend on whether a gateway answered at the moment we asked, so they are counted and left unfollowed.
What these registrations claim about themselves (names, activity, services, x402) is counted from the cards on the claims tab.
EIP-3009 authorised transfers on Base USDC · 6,060 windows read · 2026-02 to 2026-09
Not one facilitator, and not a marketplace. 19,472 different accounts broadcast the transfers we sampled, the busiest holding 5.7% of them.
Every public figure for this rail counts AuthorizationUsed events and calls them x402 settlements. EIP-3009 is a general gasless-transfer primitive that predates x402 and is used by things unrelated to it, so that count is an upper bound wearing a protocol’s name. Ours is the same upper bound, said as one.
The broadcaster is the facilitator on x402, and the only field that separates one of these from any other gasless transfer. Learning it costs a receipt per transaction, so the count comes from every window and the attribution from a capped sample, the count is the claim being checked and the attribution is what it means.
The last figure is a shape an event count cannot show: 726 wallets paying 25 destinations through one broadcaster, at an average of $0.0013 a transfer. Whether that is a service, a faucet or one operator cycling addresses is not something the chain says, so it is described rather than named, and every one of those transfers counts once in the headline that other people publish as settlement volume.
A rail is here because we read it, not because it is important. What is missing is listed too, since a list of what somebody happened to build looks identical to a list of what exists.
| Rail | Where | State |
|---|---|---|
| ERC-8004 identity registry | 19 EVM chains | read in full |
| ERC-8004 reputation registry | the same chains | read in full |
| ERC-8004 validation registry | deployed on every chain we checked | swept weekly, and it is empty |
| Masumi | Cardano | read through Koios |
| MPL Agent Registry | Solana | counted; the payload is in the Core asset |
| ERC-8004 on Solana | Solana | decoded in full |
| ERC-8004 identity registry | celo, ink, scroll, soneium | deployed and carrying single digits, left out on evidence |
| ERC-8004 identity registry | monad, hyperevm | UNMEASURED: their public nodes refuse the query |
| Virtuals ACP v1 | Base | a separate protocol, deliberately not counted with v2 |
An agent registry we have never heard of is not on this list, and nothing here can tell you how many of those there are. What the list rules out is the other reading: that these are the only ones because these are all there are.
Masumi · Cardano mainnet · registry policy and payment contract, read through Koios
Money has moved through this rail continuously for over a year, between a cast small enough to name. On ERC-8004 a registration is permanent, so the number of agents that have left is not a thing anyone can measure; here it is.
Read from the contract’s whole transaction history, not from what is sitting in it. Until today this panel reported 135 escrows, the outputs still unspent, which is a balance sheet, not a market. An escrow that settles is spent and leaves the contract, so counting only the open ones counted only the jobs that had not finished.
A Masumi registration is an NFT, so leaving is a burn and is visible. On ERC-8004 a registration is permanent and the population can only rise.
Open commitments, not settled volume, an escrow that settles is spent and leaves the contract, so this is what is currently committed rather than what has ever changed hands. All 27,948 carry a signed attestation in their datum; we record that it is present and do not verify it, because checking a signature we cannot yet tie to a registered agent’s key would be theatre. Counterparties are payment credentials in hex rather than addresses: encoding one needs a network tag and a checksum, and getting either wrong names somebody else.