The Agent Economy: Purchasing vs. Watching

The Short Version: More Listings, Fewer Purchases, and One Category Explains It

The catalog stopped shrinking. After falling from 25,529 listings in July to 14,595 in August, it grew a little, to 14,913 in September, and the number of sellers rose with it, from 1,094 to 1,127. On the supply side, things are steady.

Purchases went the other way. Paid requests across the whole catalog fell from 376,311 to 281,361 over comparable 30-day windows, a drop of 25%, and our estimate of total sales at the listed prices fell from about $40K to $32K. More listings, more sellers, a quarter fewer purchases.

That drop is not spread across the catalog. It is one category. Paid requests to web-search and scraped-data services fell from 208,728 to 80,700 — a loss of 128,028 requests, which is more than the catalog's entire net decline. Set web search aside, and the rest of the catalog grew, from 167,583 paid requests to 200,661, up 20%. One category was carrying the measured economy in August, it shrank by 61% in a month, and that shrinkage is the whole story behind the headline number. Anyone quoting a catalog-wide total for this market without that breakdown will read a downturn where the measured fact is much narrower.

Three reads of the same catalog

August's reading was that the catalog got smaller and cleaner while activity grew, which looked like dead listings being pruned rather than sellers leaving. September revises that: the pruning stopped and the catalog resumed slow growth, but the rise in activity did not hold. Taken together, the two months suggest the July-to-August jump in purchases was concentrated enough in one category to reverse within a month — a caution against reading any single month's direction as a trend. Only repeated monthly reads can show this. One snapshot cannot, and two are barely enough.

The Payment Networks Agents Use

Agents no longer pay through one system. By mid-2026 every major payment company has launched or announced a way for agents to pay. The systems differ in who holds the money and in what an outsider can see. Each status below is a dated public announcement, not a judgment of how mature the system is.

Card networks: a person lends the agent their card

Digital-dollar networks: the agent holds its own wallet

Who is this agent? Identity checks that ride on top of any network

Closed platforms: the platform holds everything

Two things follow from that last column. First, x402 is the only system whose list of services and whose payments are open for anyone to measure, which is why the rest of this report can exist at all; every other system publishes announcements, not populations. Second, every one of them has separately arrived at this report's exact question — how do you tell an agent that can spend money from a bot that only looks? The card networks answer it with identity credentials and passkey approvals; x402 answers it with per-service counts of paying wallets. The mechanisms differ; the dividing line is the same.

The ACP row deserves a closer look. The most-publicized agent checkout of 2025 was withdrawn in March 2026 for lack of conversions while its underlying standard stayed in place. That is the same gap this report measures on the x402 side, where listings and standards pile up faster than purchases.

The Three Groups, Measured at Orcpin's Own Services

Orcpin sells pay-per-request services on x402, which makes it a measuring point: every visitor is sorted into a coarse class and nothing else is kept — no addresses, no request details, no identities (see /privacy). The chart shows each group's share of Orcpin's daily visits, for the 13 days ending August 31, next to the same figure from the August edition. Shares rather than counts are deliberate: the shape of the funnel is the finding, and one seller's sales volume is private.

More than 999 of every 1,000 daily visits still involve no money. The watchers still include hundreds of daily attempts to fetch /.env and similar secret files, which is background internet noise every public server sees and has nothing to do with agents at all. The change this month is in the evaluators: their share rose from 26% to 34%. That is not because more agents asked for prices — those requests were roughly flat at about 1,650 a day — but because automated reads of the free catalog pages dropped sharply while crawler traffic stayed level. The mix changed; the actual rate of price-checking did not. Either way, money only enters the picture at the moment of payment, and paying remains by far the rarest thing on the chart. One seller is one sample; the survey below shows the same shape across the whole catalog.

Every Listing in the Catalog

On September 1 we read every listing in the CDP Bazaar, the public catalog of x402 services, along with Coinbase's own figures for how much each was used in the prior 30 days. This is the whole listed population, not a sample. The three-month table above holds the headline totals; the charts below show how those totals are distributed, and how that distribution moved since August.

Do customers come back?

This is the month's second breakdown, and it points the same way as the first. The number of repeat-use listings barely moved (1,008 to 972), but the requests they carried fell 41%, from 282,049 to 165,859 — so repeat business fell faster than the catalog as a whole. Meanwhile the one-time-use listings grew their requests 55%, from 9,509 to 14,745, on essentially the same 25 or so listings. Repeat use went from 75% of all paid requests to 59% in one month.

The one-time pattern matters because the catalog ranks services by how many different wallets have paid them. That rewards being touched once by many wallets over being used repeatedly by a few — an incentive the academic literature has described in general form (TraceRank, arXiv 2510.27554). We publish the distribution and the definitions. Whenever you see a count of paying wallets, here or anywhere, the requests-per-wallet figure belongs beside it.

What Sells, by Category

The catalog keeps confirming the same picture of demand, and this month's movement sharpens it rather than overturning it: what sells is what an agent cannot get without paying someone. That means web and social data behind paywalls, lookups of people and companies, compute and execution, and real-world goods. What does not sell is data an agent can fetch for free from any public blockchain node. The blockchain-data category now holds 5,568 listings, more than a third of the catalog, and receives 21% of paid requests — a bigger share than in August, but only because the total shrank; its own request count rose modestly, from 56,350 to 60,024. Supply is still concentrated exactly where demand is thinnest.

Two category movements are worth stating plainly because they cut against the headline. Prices and trading rose from 27,593 to 37,722 requests on a growing set of listings, and AI model calls rose from 14,336 to 17,532. The keyword rules also moved a large block: the people-and-company-lookup category fell from 1,186 listings to 519, which reflects a change in how listings describe themselves as much as a change in what exists. Read category listing counts with that caveat.

The Services That Never Take Part

Beneath the usage layer is a simpler question: does the service still answer at all? Some listed services will never see a purchase because nothing is there any more. Every day we send free test requests to roughly 500 listed services — asking for the price is free — rotating through a growing list that stood at 2,167 services on September 1, up from 1,425 in August:

How Big Is This Economy?

The one figure in this report that is a dollar total — about $32K of sales across the listed catalog in 30 days, down from about $40K in August — is the smallest of four very different ways of counting the same month. The three larger tiers (as reported on dashboards, verifiable on the public ledger, and after fake volume is removed) come from an independent March 2026 analysis and do not change from edition to edition, so they are not reprinted here. The standing reading guide — the four tiers side by side, what to take seriously, and what to take lightly — lives on the methodology page. Read this report's total against that scale, not on its own.

How We Measured

Orcpin returns counts, dates, and dollars, with the governing rules quoted. It never gives advice and never assigns a score. This report is measurement of a public system, not financial, legal, or investment guidance. See /terms for acceptable use.

Calculations

Every number on this page is either quoted from a cited source or computed from the catalog rows with one of the formulas below. The data file keeps one row per listing, carrying Coinbase's own 30-day figures: payers (quality.l30DaysUniquePayers, distinct paying wallets), calls (quality.l30DaysTotalCalls, paid requests), and the listed price (accepts[0].amount, in millionths of a dollar, so USD = amount ÷ 1,000,000). The machine-readable summary with these same buckets is generated by npm run census.

Sources