The Agent Economy: Purchasing vs. Watching

The Rails Where Agents Transact

The agent-payments landscape is no longer one protocol. By mid-2026 every major payment network has shipped or announced an agent rail; they differ in who holds the wallet and in what an outsider can measure. Status entries are dated public announcements, not judgments of maturity.

Card Rails: The User Delegates a Wallet to the Agent

Stablecoin Rails: The Agent Holds Its Own Wallet

Identity & Trust Overlays: Who Is This Agent?

Orchestration & Walled Gardens: The Platform Holds Everything

Two things follow from that last column. First, x402 is the only rail whose service population and settlement flows are open to independent measurement, which is why the rest of this report can exist at all; every other rail publishes announcements rather than populations. Second, every rail has independently converged on this report's exact question: how do you tell an economic agent from a bot that merely watches? The card networks answer it with Know-Your-Agent identity credentials, and x402 answers it with per-service payer telemetry. The mechanisms differ, but the dichotomy is the same.

The Three Populations, Measured at Orcpin

Orcpin sells x402 endpoints, which makes it a measurement instrument: every arrival is classified by coarse visitor class only, and no addresses, parameters, or identities are stored (see /privacy). The chart shows each population's share of Orcpin's daily requests over the 14 days ending August 4. Shares rather than counts are deliberate: the shape of the funnel is the finding, and purchasing volume at a single seller is proprietary.

More than 999 of every 1,000 daily requests are non-economic. The watching layer includes hundreds of daily probes for /.env and similar secrets paths, which are background internet noise that every public host sees and have no relationship to the agent economy at all. The evaluator layer is more interesting: about 1,800 times a day, something reads a paid endpoint's price and terms. A 402 challenge is free to read by design, so evaluation costs nothing, and the probe does not say whether the evaluator could have paid. Some of these callers have no wallet yet; others are funded agents whose purchasing policy rejected the offer. Either way, the economy only begins at settlement, and purchasing is the rarest event on the chart by orders of magnitude. One seller is one sample; the survey below shows the same shape across the whole listed population.

The Purchasing Layer, Surveyed

On August 5 we walked the entire CDP Bazaar catalog, reading every listed x402 resource along with Coinbase's own 30-day usage figures for each. This is the complete listed population rather than a sample.

The August walk at a glance

Compared to July

Compared with our July 16 walk of the same catalog, the listed population fell from 25,529 to 14,595, while 30-day calls rose from 313,580 to 376,311 and sellers held nearly steady (1,117 to 1,094). The share of listings with zero recorded calls collapsed from 42.7% to 0.8%. Read together, the catalog got smaller and cleaner while activity grew, which is consistent with dead listings being pruned from the index rather than with sellers leaving. Only repeated monthly walks can show this; a single snapshot cannot.

Depth Versus Breadth

The breadth pattern matters because discovery ranking on this rail follows unique-payer counts. That design rewards breadth of wallets over depth of usage, an incentive the academic literature has described in general form (TraceRank, arXiv 2510.27554). We publish the distribution and the definitions. The ratio column belongs next to any payer count you read, here or anywhere.

The Paid-Endpoint Universe, Holistically

The census keeps confirming the same demand picture: what sells is what an agent cannot get without paying someone. That means walled web and social data (Twitter search, Tavily, Exa), people and company enrichment, execution and compute, and real-world goods. What does not sell is data an agent can fetch free from any public RPC. The on-chain data category holds 5,148 listings, more than a third of the catalog, yet it receives only 15% of the calls, and its usage skews heavily toward one- and two-payer listings. Supply is concentrated exactly where demand is thinnest.

The Endpoints That Never Participate

Below the usage layer is a liveness layer: some listed endpoints will never see a purchase because they no longer answer at all. Our prober sweeps 500 endpoints daily with unpaid requests (reading a 402 challenge is free), rotating through 1,425 tracked endpoints:

How Big Is the Purchasing Economy? A Reader's Toolkit

Volume figures for agent payments arrive in tiers that differ by an order of magnitude each, and most reporting quotes the largest. The same 30 days, four ways:

The same independent analyses give the trajectory: genuine volume peaked around $5.15M/month in November 2025, bottomed around $1.19M in May 2026, and roughly doubled off that floor in June. Meanwhile the x402 Foundation launched under the Linux Foundation with every major card network as a premier member. Both facts are true at once: the purchasing economy is small, and the owners of the pipes are standardizing the meter anyway. The gap between those two facts is the industry's whole bet.

Method

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 census rows with one of the formulas below. The census keeps one row per listed resource, carrying CDP's own 30-day figures: payers (quality.l30DaysUniquePayers), calls (quality.l30DaysTotalCalls), and the listed price (accepts[0].amount, in 6-decimal micro units, so USD = amount ÷ 1,000,000). The machine-readable summary with these same buckets is generated by npm run census.

References