Someone Checked Whether AI Agents Are Paying Yet

For well over a year, the story about AI agents and money has been told mostly by the people building the rails. A protocol launches, a network announces a product, a transaction count goes up, and the conclusion people draw is that agents are starting to pay for things. I’ve written in that neighbourhood myself. What was missing was someone going to the record and checking who was making the payments.

TRM Labs, a blockchain intelligence firm, has now done that for x402, and very little of the activity it looked at can be shown to come from agents. I think that finding deserves more attention than any launch this month, partly for what it says about agent payments and partly for what it says about how this industry decides what’s true.

What TRM measured

x402 is a payment standard Coinbase launched in 2025. It takes the HTTP status code 402, “Payment Required,” and turns it into a way for software to pay per request without a checkout page. A client asks for a resource, the server answers with a price, the client pays, and access follows. It was designed with agents in mind.

TRM looked at x402 activity on Base, Solana and Polygon from May 2025 onward: roughly $52.7 million in value across 198.9 million settlement transactions. It then screened out what didn’t look like commerce, removing addresses paying themselves, large flows dominated by one or two payers, and sellers with fewer than 10 distinct buyers. That left $25.62 million of likely commerce. To that, TRM applied two tests for whether a payment looked agentic, one permissive and one strict. The permissive test counts anything that plausibly could be an agent. The stricter evidence it looks for includes sustained activity over multiple months, registration in a public agent registry, or payments to multiple sellers. Depending on which test you use, between 0.6% and 7.5% of the screened commerce appears to be agentic.

One more number is worth having in view. Of $52.68 million in settled value, $52.47 million, or 99.6%, was in USDC. TRM published its analysis on September 9, and PYMNTS reported it on September 15.

What the rest of it is, and isn’t

The easy summary is that more than 92% of x402 payments come from scripts rather than agents. PYMNTS’s own headline says most x402 payments aren’t from AI agents. That’s a little stronger than what TRM found, and the difference matters for the argument I want to make.

TRM’s point is that the chain can’t tell the two apart. In its own words: “x402 was designed for agentic commerce, but nothing in the protocol requires an agent. Anyone who writes a script can drive the same 402 sequence, and it leaves an identical onchain record.” A scheduled job paying an API every few minutes and an AI agent deciding to pay the same API produce the same transaction. So the remainder isn’t a measured population of scripts. It’s activity TRM couldn’t show to be agentic. Some of it is almost certainly ordinary automation. Some of it could be agents that don’t look like agents, and TRM says so itself: part of its screening “may understate the space,” because many agents today could be single-purpose, paying one service repeatedly, and the test would read that as a script.

The accurate reading, then, is that the agentic share of x402 commerce is small and uncertain in both directions, and even the permissive end of the range is small. By my own arithmetic, 7.5% of $25.62 million is under $2 million, across all the activity since May 2025 on three chains. That’s the most generous estimate of what could plausibly be agents.

The narrative was a self-report

The main argument on this site is that an agent’s self-report is not evidence. An agent describes a finished job in the same confident tone whether the job went well or not, so the check has to come from a record the agent didn’t write. I think the industry story about agent payments has had the same shape, one level up.

As far as I can tell, most of what people repeated about agent payments over the past year came from parties with a stake in it being true: protocol teams, networks, wallet providers, and commentators, including me, who find the idea interesting. None of that is dishonest. It’s the normal way a new category gets talked about. But it is a self-report. The figure most often attached to it was transaction volume, and transaction volume measures the rail. It tells you how many payments crossed x402. It tells you nothing about what made them, and what made them was the claim being implied.

What makes TRM’s work useful is that it went to the independent record and asked the question the narrative skipped. The answer also contained a second finding that I’d put next to the percentage: the independent record, on its own, couldn’t fully answer the question. A public chain shows that a payment happened, between which addresses, for how much. It doesn’t show whether software with any judgment was behind it. TRM had to build tests on top of the record, and be explicit about where those tests could be wrong, to get as far as a range.

My working days are in payment orchestration, and the habit I trust most from that world is reconciliation. What a system says happened gets checked against a separate record of what did happen, routinely, and not only when something looks off. The agent-payments conversation has mostly skipped that step, and it isn’t the only part of this space that has. When I wrote that agentic checkout is being built faster than anyone trusts it, the gap was between the infrastructure and consumers’ willingness to use it. PYMNTS frames this one as infrastructure that may be developing faster than the agents themselves. To me those look like the same pattern seen from two sides.

Where this leaves my own stablecoin argument

In Most Agent Payments Won’t Run on Cards, I argued that agents paying other agents, and paying for the services they use, is a larger category than agents shopping at merchants that take cards, and that I expect most of it to move over stablecoin rails rather than card rails. That’s a view about where the money goes when agents pay, and about direction. It wasn’t a claim that agent payments were already large, and TRM’s data doesn’t say much about the rail question either way.

I want to be careful with the one number that looks like it helps me. 99.6% of x402 value settled in USDC, but as PYMNTS puts it, x402 is a stablecoin payment system, and nearly every merchant TRM identified specified USDC as its preferred payment asset. That share reflects the protocol’s design. I’m not going to count it as evidence that stablecoin rails are winning anything.

I should also be precise about what that piece did and didn’t say. It called agent-to-agent payments “the larger version,” meaning over time. It didn’t put a figure on how much of that was happening today, because I didn’t have one. A reader could reasonably have come away thinking the activity was already well underway, and the piece did nothing to correct that impression. Now there’s a figure, and it says the category is mostly still ahead of us. That doesn’t change my view on the rails. It does mean I’d be more careful about timing than I was.

The other half of that piece is where TRM’s work lines up most closely with what I argued. I said a wallet’s public history is a list of transfers that doesn’t tell you which agent a payment belongs to, and that the check on agent spend has to come from reading the chain independently rather than trusting the agent’s own account. That was an argument about one operator watching its own agents. TRM has run the same check for the market as a whole, and hit a more basic version of the same problem: the record doesn’t say whether an agent made the payment at all. Its recommendations follow from that. It calls for “accurate registration, counterparty reputation an agent can check on its own, and monitoring built for volume rather than value,” and concludes that “agentic commerce will need agentic compliance.” The analysis also notes, as PYMNTS reports, that existing agent registries let people declare ownership of agent addresses, but participation is voluntary and uncommon. The case for monitoring by volume makes sense to me given the shape of what x402 carries, which is a very large number of small payments, where a check keyed to dollar amounts would rarely have anything to catch.

Where this argument stops

This is one protocol. Agents that pay by card, through products like Mastercard’s Agent Pay, or through stablecoin mechanisms outside x402, aren’t in this data, so the finding is about x402 rather than about agent payments everywhere. The window starts in May 2025, which is early for any payment standard, and rails usually arrive before the traffic that justifies them. A small agentic share now is compatible with a large one later, and I’d be making the same mistake in reverse if I treated this as proof that agent payments aren’t coming.

The measurement has limits too, and TRM names them. Classifying a transaction as agentic from on-chain behaviour is inference, and part of the screening may miss single-purpose agents. Better registration would narrow the range, which is part of why TRM asks for it.

What I take from it is narrower than either the hype or its opposite. When someone cites x402 transaction counts, or any payment volume, as evidence that agents are paying, the question to ask is what share of it can be shown to come from agents, and how they know. The narrative mostly skipped that question. TRM asked it, and the answer is a small number with a wide range around it, which I find more useful than the confident story that came before.