TRM Just Audited x402. Between 0.6 and 7.5 Percent of the Payments Are Coming From Agents.

TRM Labs published a chain-forensics look at x402 this week. It settled the argument that a lot of people in agent commerce had been avoiding: is x402 an agent-payments protocol, or is it a payments protocol that some agents use.

The receipts, since May 2025 across Base, Solana, and Polygon: $52.7 million in raw settled value across 198.9 million transactions, processed by the facilitators that broadcast the majority of x402 traffic. USDC ate 99.6 percent of the settled dollars ($52.47M of $52.68M). The rest is rounding on a stablecoin protocol.

TRM stripped out addresses paying themselves, flows dominated by one or two payers, and sellers with fewer than 10 distinct buyers. That collapsed the raw $52.7M into $25.62M of likely commerce. Of that $25.62M, the fraction that shows the signature of an autonomous agent lands between 0.6 percent and 7.5 percent, depending on which test you run.

The Two Tests

TRM ran two heuristics.

Test one: was this a facilitator-broadcast payment with variable amounts averaging below one dollar. This is the "many small purchases at different prices" signature of an agent shopping across endpoints. Result: 0.6 percent of screened commerce.

Test two: did this address sustain activity across multiple months, register in a public agent directory, or pay multiple sellers with different amounts. This is the long-lived agent signature. Result: 7.5 percent of screened commerce.

Neither test is proof. TRM concedes both directions of error: a single-purpose agent that hits one endpoint at the same price for a year looks exactly like a script, and a cron job on Base that pays 40 different sellers looks exactly like an agent. Agency is not visible on-chain. What TRM measured is agent-shaped behavior, which is the closest thing to a floor and a ceiling anyone has published on this protocol.

The Numbers, Ranked

Take the generous end of TRM's range and extrapolate against the Bitquery August number. If 7.5 percent holds across the newer, larger monthly volume, agent payments on x402 land near $195 million a month against a $2.6 billion protocol total. If the pessimistic 0.6 percent holds, the number is closer to $15.6 million a month. Both numbers are real, and neither is the number a founder wants on a slide when raising a Series A on x402 rails.

Where the Volume Is Actually Coming From

TRM's screening framework, read forward, tells you what the discarded $27M actually was.

Self-payments and back-and-forth transfers between wallets the same operator controls. Nobody is buying anything, someone is looping. Flows dominated by one or two payers, which are single automated buyers hitting a small pool of sellers, often the same one. That is a script. Sellers with fewer than 10 distinct buyers, which are private test rigs, proofs-of-concept an agent-commerce startup is running against its own frontend, or a small closed-loop deployment.

Then the residual $25.6M of genuine commerce gets sliced. The 0.6 to 7.5 percent that reads as agentic is the top slice. The other 92.5 to 99.4 percent is people, running scripts, paying for APIs and content the same way people have paid for APIs and content since 2010, with a new settlement layer underneath. That is not a scandal. That is the story of the settlement layer maturing ahead of the customer.

The AFTA Read

TensorFeed operates the AFTA rail, so we watch this data more carefully than most. Agent-fair-trade certification is priced against the assumption that the counterparty on the other side of a paid endpoint is, in fact, an agent. If it is not, most of AFTA's guarantees still fire (5xx no-charge, breaker, schema-fail refund, Ed25519 signed receipt), and none of them stop firing. But the verified agent traffic narrative sits on a smaller pile of receipts than the wires imply. Three consequences worth naming.

One, the pricing floor for agent payments is set by settlement mechanics, not by agent adoption. USDC on Base at 99.6 percent tells you the cost curve is stablecoin infrastructure and facilitator margin. The number of literal agents in the loop does not move that floor down until it moves the volume up by an order of magnitude.

Two, verified feed matters more, not less. When two thirds of "agentic" activity turns out to be a cron job on a rented VPS, the ability to say cryptographically who is on the other side of a payment becomes the interesting product surface. If most x402 traffic is scripts, then the receipts, registries, and signed manifests that separate a real agent from a well-configured script become the actual moat. That is the AFTA thesis restated in TRM's methodology. We wrote it up in April when we picked USDC on Base as the AFTA settlement layer, and the TRM audit is the first outside dataset that makes the receipts case for us.

Three, the agent economy GDP number the ecosystem has been quoting is a ceiling, not a floor. Andrew Peek at TRM said it plainly: agency is not visible on-chain. Any agentic-payments TAM that starts from settled dollars is measuring settled dollars, not agents. The TAM has to be measured somewhere else, and that somewhere else is not on the ledger.

Counterreads

Three counterreads deserve a fair hearing.

First, TRM's methodology probably undercounts. A well-designed agent hitting one endpoint over and over is indistinguishable from a script, and every serious agent framework right now runs the same call in a loop. The 7.5 percent ceiling is not the ceiling on agent activity, it is the ceiling on agent-shaped activity that TRM's two heuristics can detect. The true agentic fraction is somewhere higher, and the honest response is that we do not know how much higher, and neither does TRM.

Second, x402 is 16 months old. Every payment protocol looks like this in its first phase (scripts and dev-loop traffic ahead of production customers), and the agents are not yet the payers reading is a snapshot, not a trajectory. Card networks in 1994 were mostly phone order and mail order, not internet commerce. The internet did not arrive because the tape said it had; it arrived when the infrastructure could carry it.

Third, the surface area is wider than x402. Google's AP2, ACP, MPP, and the payments extensions Mastercard and Visa are building on their own rails all count in the agent-payments column, and TRM only screened one protocol. A composite agent-payments-across-rails number could look different from a Base-only x402 number. The composite has not been measured yet either.

All three reads are defensible, and each of them adjusts the finding rather than deleting it. The number a founder pitches on Monday is still not the number TRM published on Tuesday.

Our Take

The interesting fact is not the 0.6 percent, and it is not the 7.5 percent. It is that the audit exists at all.

For 16 months, agent-payments discourse has priced its own thesis by pointing at x402 volume growth as proof that agents are paying. Nobody in that discourse controlled for whether the counterparties on the other side of those payments were agents. The measurement did not exist because the measurement was inconvenient, and the ecosystem raised on volume because volume was countable. TRM published the first serious "who is actually on the other side of this" analysis, and the answer, at the honest ceiling, is less than one dollar in twelve.

For the next quarter, the useful posture is not defensive. It is to build the receipts that make an agent counterparty legible on-chain in the first place. That is the product opportunity TRM's paper opens, and it is one AFTA has been arguing for since the whitepaper shipped. The x402 rail is real, the USDC settlement floor is real (see our batch settlement piece from May for the mechanics), and the agent share is small and improvable. All three statements are consistent, and the third one is where the next 90 days of product work should sit.

Three signposts for the next 60 days. One, whether Coinbase, Circle, or the x402 Foundation publish a first-party agent-share metric (a monthly number the ecosystem can cite against TRM's snapshot). Two, whether a public agent registry gains meaningful adoption, so an on-chain address can prove agent identity as a first-class field on the payment (we shipped one piece of this with the verifier MCP in June). Three, whether the next major agent-commerce startup raises on TAM figures that pass TRM's screen (variable amounts, multiple sellers, sustained activity), rather than on gross settled volume. If two of the three fire, the agents are the payers story stops being aspirational inside a quarter. If none fire, next quarter's headline number will be bigger and less informative, and TRM will get to publish this piece again.