95,882 Agent Identities on Base, 1,198 Ever Paid Anyone

There are 95,882 AI-agent identities registered on Base under ERC-8004. The number that have ever paid anyone, from a wallet you can prove is theirs, is 1,198 — one in eighty.

That figure comes from Agentic Finance Graph, which describes itself as an independent public ledger of agent money and says it moved from sampling the ERC-8004 registry to reading all of it. The count is the group's own, derived from its own join method, and it is worth reading the method before quoting the headline.

The join nobody publishes

Registries count identities. Payment rails count transfers. The link between them — this token spends from this wallet — is what every "agent economy" dashboard draws as an arrow and nobody publishes evidence for.

The group's answer is a binding record that is born, dies, and can be superseded rather than edited in place. When an identity NFT is resold, the binding closes at that block and the new holder has to earn a new one. Binding methods are ranked, never averaged: a wallet signing a challenge over the agent ID plus a nonce scores 0.95; an EIP-7702 or session policy embedding the agent ID scores 0.90; a registration that declares a wallet later observed spending scores 0.70; ownerOf(tokenId) as the observed spender scores 0.55. Name similarity and clustering are never a binding.

The exclusions are the story

A stablecoin transfer counts only if the sender has a binding valid at that block, the block is at or after the identity's mint, and the receipt does not show money arriving and leaving the same wallet in one transaction — a router hop, not a purchase.

Applying those rules to every dollar that left a bound wallet, as of 27 September 2026, the group reports: $8.8M in counted payments, $72.2M excluded as spending that happened before the identity existed, and $17.8M excluded as routing hops, out of $98.8M total. A naive total is roughly 11 times the filtered one.

One wallet explains why the date rule exists. The post cites wallet #63877, which registered an identity on 18 August 2026 after having already moved $34.6M in stablecoins. Without that rule, a single identity would have added four times the entire counted total to "agent spending".

Where the counted money goes

The composition undercuts the commerce narrative more than the headline ratio does. Of the counted $8.8M, the group attributes roughly 60% to lending and vault adapters — agents parking their own cash at yield, with the single largest recipient a verified contract that wraps Aave v3 and ERC-4626 deposits. Another 18% goes to bridges, 4.5% to escrowed work, and 1.5% to swaps.

Two more numbers set the scale. The largest wallet accounts for 28% of the total, and the median payment is $48. Only 5% of counted transfers are x402 micropayments, which suggests the wallets buying APIs over HTTP and the identities registered under ERC-8004 are still mostly different populations.

What it does not claim

The group is explicit about its own limits, which is the main reason to take the numbers seriously. A binding is evidence of control, not of operation — the chain does not say whether a program or a person signs. A label describes what a contract does, never who runs it. A counted payment is a transfer, not proof that anything was bought.

If you are building on agent payment rails, the practical read is that registered identity and economic activity are close to uncorrelated right now, and any vendor citing registry counts as demand is citing mints. The group publishes per-agent records through its API, so the claim is checkable against a specific identity rather than taken whole.