Counting the agent economy

Between 5 September and 5 October 2026 I measured the x402 Bazaar twice. Once out of the catalogue itself, and once against Base. The catalogue came to 35,000 rows and 34,010 distinct entries.

On the chain side I took every USDC inflow to the 1,394 payTo addresses the catalogue names. Then I threw most of them away. What stayed were the inflows whose amount matched a price that address itself advertises. A salary does not arrive at a listed price.

The counters

Every entry carries its own call count, and the first thing to establish is who produced it. Neither published version of the extension specification defines the field at all, so a seller has no channel to declare one, and neither of our own two services ever sends it. The facilitator is counting.

The chain says the same thing, less abstractly. The largest entry in the catalogue reports 354,713 calls at two cents. Its payTo address received 354,819 USDC transfers in the same window, worth 7,096.38 USDC. Both the count and the amount sit three hundredths of a percent from what the entry claims about itself.

The instrument was checked twice over. Two unrelated RPC providers read the densest day of the window. Both returned 246.140000 USDC across 12,307 transfers, to the microcent and the single transfer.

Two markets

Most of the confusing numbers here come from one habit: treating agent payments as a single market. There are two. On card rails a person grants a mandate in advance and the agent then buys inside it, which is the shape Stripe's Agentic Commerce Protocol, Visa Intelligent Commerce, Mastercard Agent Pay and Google's AP2 all have in common, and because a human signed something first, every figure from that side describes a flow that a person authorised before the agent moved — which is a different quantity from the one the second market reports, and the two are routinely added together anyway.

What that side has published is thin to the point of a single word. Visa said in December 2025 that hundreds of agent-initiated transactions had been completed. The figure has not moved since.

Mastercard has published nothing at all, and told its first-quarter call that volumes were at an early stage.

Stripe's announcements name no orders and no GMV. Google has never published a figure for AP2 in any form.

The one large measured figure on that side is Chinese. Alipay reported more than 120 million AI Pay transactions in a single week in February 2026. It is a count with no amount attached, and it covers a wider surface than chatbot checkout, since voice purchases on smart glasses sit inside it.

The merchants say more than the networks do. Etsy put agentic traffic at under one percent of its total on its second-quarter call. OpenAI withdrew Instant Checkout on 24 March 2026, six months after launching it.

The second market pays per API call. No human approves the individual transaction, and x402 is the only rail there with a usable record.

Visa and Artemis put the lifetime figure at 15.0 million dollars across 109.6 million transactions once wash and test activity is removed, and the shape of what that exercise threw away says as much as the figure it left: the cleaning took out 88.9 percent of the dollars while taking out only 38.5 percent of the transactions, so the traffic that failed the test cost far more per call than the traffic that passed it, which fits a discarded volume of farming.

TRM Labs filtered further. It put the genuinely agentic share between 0.6 and 7.5 percent, or five to eleven thousand dollars a month.

Stripe and Tempo's Machine Payments Protocol has recorded about 25,000 dollars on the same kind of accounting.

Six figures in circulation

The counter on x402.org reports 75.41 million transactions and 24.24 million dollars under a heading that says Last 30 Days. The same four values appear in Wayback snapshots from 1 August, 1 September and 1 October 2026, and in the live page today. That makes the window ten months wide and the volume figure twenty-six times the measured one.

A claim that MPP processed fourteen million dollars in sixty days cites a Stripe page that contains no dollar figure anywhere. The supportable number is 25,000.

Mastercard's line about nearly all its cards being enabled for Agent Pay describes credential provisioning, not use, and the same earnings call put volumes at an early stage.

Nevermined's homepage reports paid calls and dollars settled today. Both values are hardcoded in the raw HTML.

Virtuals reports an agent GDP of 481.8 million dollars. Its own glossary defines that as pass-through notional, so a five-thousand-dollar trade carrying a ten-dollar fee counts as 5,010 dollars of agent GDP.

Seller counts for x402 inside one quarter run 100,000-plus, 22,000, about 5,300, 3,892 and 1,387. The spread is twenty-six to seventy-two times, and every one of the five is somebody's published figure.

The denominator

One sentence came out of the month with a denominator, a chain measurement and a date attached to it: the endpoints listed in the CDP x402 Bazaar turn over roughly 18,000 USDC in thirty days, measured as price-matched USDC inflows on Base, as of 5 October 2026.

The counters, multiplied by each entry's list price, give 16,760.42. The chain gives 18,097.82. Eight percent higher, and in the direction the mechanism predicts, because a seller only enters the catalogue by opting in and the counter is therefore a floor rather than a ceiling.

The gaps

The number covers USDC on Base. That leaves out Solana, BNB Smart Chain, Polygon, Stellar and XRPL, and it covers one facilitator's catalogue and not the protocol, so of the 54,201 resources standing in the eight publicly readable x402 catalogues I could crawl, 36 percent are outside the CDP Bazaar and invisible to every figure above.

The third gap is the widest. It covers catalogued endpoints, not a facilitator's traffic: CDP settled about 531,000 dollars across 1.92 million settlements in the same thirty days, which puts the catalogued entries at 3.4 percent of that value.

Two of the gaps will not close with better research. Settling without a facilitator is a supported path in the specification, and an ordinary EIP-3009 transfer carries no facilitator identity. That traffic sits in nobody's denominator, and an open issue in the specification repository says the same about what an outside observer can determine from chain data alone.

No test-and-wash adjustment was applied to my own thirty days either. The one comparison available for that correction removed 88.9 percent of the dollars.

Visa reports on 27 October and Mastercard on 29 October. Either call could put the first measured figure on the card-rail side, and the Artemis dataset has not been touched since 21 April. I will re-run the chain measurement after both dates and publish what moved.

Evidence