Can you prove an AI agent made a payment? Not from the chain alone.

Why a cron job and an agent leave the same trace

A settled x402 payment on Base records a payer address, a payee address, an amount, a token, a block height, a timestamp and a signature. It does not record intent, session, model, or whether a human pressed a button.

Schedule a script to hit a paid endpoint every sixty seconds and the chain shows a repeated transfer between two addresses. Let an autonomous agent decide to buy a dataset and the chain shows the same fields. Identical records, identical finality, identical explorer page. The chain settles payments. It does not attribute them.

x402 does not close the gap. According to the x402 protocol documentation, the protocol charges no fee of its own and payers cover network fees only. That keeps settlement cheap and keeps the payment envelope thin: there is no agent identifier field, no attestation of who authorised the transfer. Blockchain-agnostic support for EVM chains and Solana widens reach without adding provenance.

So when you read a number for "agentic payments", the first question is what the author did to separate an agent from a loop.

What TRM Labs actually screened for

TRM Labs published the most careful public attempt. In Who's Actually Paying? Measuring AI Agent Payments Onchain, they examined roughly $52.7 million across 198.9 million x402 settlements on Base, Solana and Polygon since May 2025.

Then they screened, and the steps are described well enough to repeat:

1. Remove self-transfers and internal wallet movement where one entity sits on both sides of the transfer.
2. Remove wash-like patterns where funds circulate back to a common source.
3. Set aside round-number stress tests and load-testing traffic that never varies in amount or cadence.
4. Keep the remainder as commerce: value that looks like someone paying for something.

That leaves $25.62 million. Less than half the headline figure survives the screens, and of the screened total, 0.6% to 7.5% appears agentic.

The band is wide on purpose. It comes from judgement calls about which patterns count. TRM states the agent figure as roughly $5,000 to $11,000 a month, not as a lifetime dollar range. Anyone quoting 0.6% to 7.5% as a precise share is reading more into the method than the method supports.

One number in the set is clean. Of $52.68 million, $52.47 million settled in USDC, which is 99.6%.

The other counts, and what they leave out

Chainalysis adds two useful details. Payments of $1 or more rose from 49% of value in early 2025 to 95% by early 2026, which suggests value concentrating in fewer, larger calls. And PING, a pay-to-mint experiment, lifted transactions by over 10,000% in a single week. One campaign can dominate a transaction count, which is exactly why raw totals need screening before anyone calls them adoption.

The direction of travel has not been one-way either. BlockEden.xyz reports that daily x402 transactions fell from about 731,000 in December 2025 to about 57,000 in February 2026, a decline of over 92%. CoinDesk's read on the demand side is blunt: the protocol fixes micropayments, and the volume is not there yet.

What a seller can log that the chain cannot

Your edge sees the request. The chain sees the receipt. Record at the request layer and you have evidence no explorer can give you:

• The user agent string and any framework headers identifying an agent runtime.
• Whether the X-PAYMENT header was present before the first byte, or added after a 402 challenge.
• Session shape: a single fetch against repeated calls with varying parameters. Cron looks regular. Agents look bursty.
• Wallet-to-session mapping, where one payer address shows up across many independent sessions from the same client fingerprint.
• Whether a human fetched the same URL afterwards, which usually means a person was reviewing what an agent bought.

Cloudflare already ships tooling in this direction. The Monetization Gateway, published 1 July 2026, lets customers charge for any protected asset, including pages, datasets, APIs and MCP tools, settled in stablecoins over x402. Pay Per Crawl puts Cloudflare in the merchant-of-record seat for crawler access using HTTP 402. AI Crawl Control lets publishers sort crawlers by purpose instead of blocking them wholesale. If you are deciding whether to charge at all, that trade-off has its own post.

None of this attributes an agent on-chain. All of it attributes at the edge, where the evidence actually lives. Per W3Techs, Cloudflare sits in front of 25.8% of all websites as of the September 2026 survey, so edge logs are the majority-case instrument for this problem. Building the payment side without the logging side leaves you with revenue and no provenance. The two halves belong together.

Why every agent-economy figure inherits this weakness

Attribution happens at the wrong layer. Every published agent-payment statistic is a settlement statistic with an assumption bolted on, and the assumption is rarely stated.

A transaction is a fact. The identity of the payer is an inference, and the inference is where the arguments live.

The incentive runs one way. Coinbase's own cumulative count in late April 2026 came to roughly 165 million transactions, about $50 million of volume and 69,000 active agents, published the week Agent.market launched. That timing does not make the number wrong. It tells you what the number was built to support.

Standardisation may help eventually. The Linux Foundation announced the x402 Foundation on 2 April 2026 at the MCP Dev Summit North America in New York, with Coinbase contributing the protocol. The operational launch followed on 14 July 2026 with 40 member organisations and 17 premier members: Adyen, AWS, American Express, Circle, Cloudflare, Coinbase, Fiserv, Google, Mastercard, Monad Foundation, MoonPay, Ripple, Shopify, Solana Foundation, Stellar Development Foundation, Stripe and Visa.

Governance is not provenance. None of those memberships creates an agent identity field on a settlement.

If you sell to agents, log the request layer, publish your screening method, and state the range rather than a point estimate. Kunzum's own scoring scripts and raw responses sit in the published dataset under CC BY 4.0, with no gate, so you can check how the screening worked here instead of taking the summary on faith.

The honest position on any single on-chain payment: nobody can tell from the available data whether an agent made it. The transaction is provable. The initiator is not.