The tape, with receipts.

Float & treasuries

The pulse funnel — every rung, measured

99,002 names. 7,710 that pay.

This is the whole argument in one picture. Each rung is a harder question than the one above it, and every bar is something we measured rather than modelled. The first five come from a uniform random sample of the registry, so they are directly comparable. The last one is counted, not sampled — and it is a floor.

The bars are linear on purpose. A logarithmic scale would turn the drop from declares a service to actually answers into a gentle step, and that drop is the finding. Rungs L1–L5 are shares of a uniform random sample of 1,024 registrations, so the percentages are comparable to each other. The paying rung is different in kind: it counts agents whose wallet we have watched send money, against every registration ever minted — a floor, because we have not scanned every wallet. How each rung is defined.

L7 — observed payment · the rank that cannot be faked

We watched these wallets pay somebody.

Every rank below L7 can be inferred from a registry: a name exists, a file parses, an endpoint answers. L7 cannot. Somebody has to watch a wallet send money to a different address and record the transaction. This is that record — outbound USDC on Base, self-transfers excluded, because moving your own money between your own wallets is not paying anyone.

x402 — the rail, measured on our own payments

Everyone quotes x402 volume. Almost none of it reaches an agent we can see.

We already watch every outbound payment from the agent wallets we have bound. Once you know which addresses actually receive x402 payments — harvested from a public index, with the citation kept — you can ask a question nobody publishes: of the machine payments we observed, how many went over the x402 rail? The answer is small, and the payments are tiny. Treat it as a floor: the recipient set is incomplete by construction.

Payments — how machine money actually settles

Identity & liveness — measured by us

What we can prove about registrations.

Every figure below is our own observation, refreshed by the collector — not a number borrowed from someone else's crawl. Machine-readable with producer and definition at /api/state.

The join — registration to wallet to money

We followed the money ourselves.

For every registration in our sample we read its owner on-chain, read any payment wallet it declares, then asked the USDC contract what those wallets actually hold. No estimate, no extrapolation — this is the sample, counted.

Reputation — a signal, never a rank

Agents are rating each other, on chain.

The ERC-8004 Reputation registry carries live agent-to-agent feedback: a rater, an agent, two short tags, and a link to an off-chain scorecard. We read it straight from the registry's own logs. It sits beside the liveness ladder and never inside it — feedback is somebody's opinion, and an opinion can be bought as easily as it can be earned. What we watched a wallet do is a different kind of fact.

The demand side — where the money actually lands

Agents are not paying people. They are paying contracts.

A registry tells you who exists. A payment scan tells you who spends. Neither tells you where the money goes, so we asked: rank every destination by how many different agents pay it. Distinct payers, not volume — volume is dominated by one large actor, while the number of separate agents choosing one destination cannot be faked by a single whale. The answer is a short list, and it is mostly code.

Bridges — the other side of an exit · added 30 September 2026

Money that left Base, followed to where it arrived.

A payment into a bridge is where a Base ledger normally stops. Each bridge keeps the other half in its own public status record, so we ask it: for every counted agent payment into Across, LI.FI, Relay, deBridge or Circle's CCTP, which chain the money reached and which address received it. The Base side is our receipt-checked payment; the destination side is the bridge's record. None of it is added to a Base figure, and moving money is never called spending.

Who mints — the headline, taken apart

Every "how many AI agents" number is a mint count.

So we asked who does the minting. One address is responsible for an eighth of the entire Base registry, the ten busiest for nearly a third, and almost a quarter of all registrations arrived in bursts of twenty-five or more from a single address inside one hour. None of that is fraud — bulk registration is a legitimate thing to do. It just means the headline counts names issued by scripts, and reading it as a population of working agents is a category error.

The resale market — identities change hands

An agent registration is an NFT. Nearly a quarter have been sold.

Every headline about how many AI agents exist counts registrations. But a registration is a transferable token, and we counted the transfers straight from the registry: a large share have already moved to a different owner. That does not make them fake — it makes a registration count a count of names issued, not of agents, and certainly not of distinct operators.

The keys — who actually signs

One in six agent wallets has handed over its keys.

EIP-7702 lets a plain wallet point at contract code. After that, the wallet does whatever that contract says when it signs. We read the code of every bound wallet directly — the delegation leaves an exact 23-byte fingerprint on chain, so this is counted, not estimated. Nobody else is publishing it, and it is the largest unpriced risk in agentic finance.

What we found — every discovery, kept

The graph doesn't just count. It notices.

A number tells you the total moved. A finding tells you which registration paid, when it first did, and where to check the transaction yourself. Each one below came out of our own scan; the strongest are promoted to markers on the chart at the top of this page. Everything here is also served at /api/findings.

— · JSON API · how the ranks work

Independent research — cited, not ours

Work by other people that we have not reproduced. Kept separate from our own numbers on purpose. Where our measurement covers the same ground, we say so.

Context — mood & market backdrop, kept small on purpose

Forecasts — quarantined on purpose

Gartner projects $15T of agent-intermediated purchases by 2028; McKinsey sees $3–5T of retail agentic commerce by 2030. These are forecasts — narrative ceilings, not measurements. They never appear in a live tile on this site, and if you see them quoted as current AUM anywhere else, close the tab.

The metrics this page said were coming

When this page opened in early September, these six were a promise. Four are measured now and two are partly built. Every value below comes from the same tape as the rest of the page.

• Agent Treasury Utilization live: what the wallets bound to ranked agents hold, priced. Cash (stablecoins) 10.01%, ETH 51.66%, DeFi positions net of debt 0.01%, other tokens 38.32%, across 900 wallets. Tokenised real-world assets are not split out yet; one would sit under tokens or DeFi. /api/treasury · definition
• 7702-delegated share vs god-key share live: counted agent dollars by who can sign for the paying wallet. A single plain key (the god-key case) 26.16%, a key delegated to code under EIP-7702 9.69%, a contract account 64.15%. Of the bound wallets examined, 15.73% carry a 7702 delegation. definition · Why it matters →
• Vanity ratio live: x402scan publishes a normalised 24-hour volume beside its raw one, and the normalised figure is 0.19% of the raw. The gap is what a handful of outsized transfers add. What the sellers listed in the Bazaar actually receive, and how much of it registered agents pay, we measure ourselves: /api/x402flow.
• Raw dumps CSV live: the whole archive as a flat file, one row per point and one column per metric, at /api/series?format=csv (add &keys=a,b for chosen columns). Parquet on request.
• Liveness funnel by chain partly live: on Base, rungs L1 to L5 from a uniform sample of 1,024 registrations, re-drawn daily. On Solana, 43.2% of 1,169 agent cards probed are valid. On BNB Smart Chain, Ethereum, Arbitrum, OP Mainnet and Polygon, registrations are counted and their owners' payments read, but the cards are not probed yet; that, and a weekly roll-up per chain, come next. The coverage verifier runs: 40 data checks, none failing on data health, and the BNB payment classes publish only once 98% of receipts are read.
• The agent multiplier half built: 3.07 Base registrations per minting address, the largest single minter holding 10.38% of all of them. An address is not a person: a platform minting for its users pushes the figure up, one operator spread over many addresses pushes it down. The second half, transfers per agent per day against a baseline of ordinary wallets, needs that baseline sampled first.

Want a metric added, or have labelled wallets to contribute? agenticfinancegraph@proton.me · @AgenticGraph.