On 2026-10-14, Chainalysis reported that AI agents had made more than 138 million x402 payments on Base by the third quarter of 2026. Adding Solana and other chains, the ecosystem was above 165 million transactions and about $50 million in cumulative volume. That is a genuine milestone for machine-to-machine payments, and it is the kind of number that gets quoted without its denominator.
We run a small x402 marketplace: 1,709 listed endpoints, 1,636 of them with a free-trial allowance. We read the same protocol from the other side of the wire. This post compares the headline to what our own ledger shows, because the two are not contradictory — they answer different questions.
Several details from the Chainalysis coverage matter more than the total:
There is also a caution worth repeating: rival analytics firm TRM has argued that much apparent agent activity may be scripts or self-dealing rather than autonomous agents, and Chainalysis itself says its figures cover identifiable x402 activity, not the whole market. Those caveats do not cancel the milestone. They tell you what the number is for.
Our own /api/stats endpoint publishes a deliberately narrower set of figures. Measured on 2026-10-20:
| field | value | what it counts |
|---|---|---|
| services | 1,709 | listed endpoints |
| endpointsWithTrials | 1,636 | endpoints with a trial allowance |
| paidCalls7d | 17 | settled external paid calls, 7 days |
| agentPayingWallets7d | 5 | distinct signed wallets that paid, 7 days |
| realAdoptingAgents7d | 2 | distinct self-identified agent ids, 7 days |
Seventeen paid calls in a week is not a growth story, and we do not present it as one. But the composition is the interesting part. Our internal paying-wallet provenance check reads the user-agent each payer sends and annotates the wallet list. In the 7-day window ending 2026-10-19, the credit ledger held six distinct paying wallets. Five of them self-identify as directory or verification services:
market402.com — paid delivery probenohumans.directory — paid censusprobe402.com — settle legsettled.tools — pays once and records deliveryannalist.dev — independent x402 checksThe sixth had no paid delivery row in the window. A user-agent is a claim, not a fact, and our own check says so. But the picture is consistent: at our size, most paying traffic is the verification layer of the ecosystem, not end customers. That is not necessarily bad news. Someone has to test whether a rail settles before anyone trusts it with real volume.
One report carried in the Cloudflare coverage cited x402 settlement volume down 93% since early 2026. We have not independently measured that, and it may describe a different slice of the market than Chainalysis's count. Still, it is not paradoxical for transaction count to rise while dollar volume falls. If most new transactions are one-cent inference calls or paid probes, the count goes up and the average ticket goes down. Both statements can be true at once.
The practical consequence is that a marketplace should be designed for the one-cent case first: minimum price floor, per-call metering, no top-up threshold that a $0.01 buyer cannot clear, and a trial path that lets an agent verify the response before it signs anything. That is what we build for, and the market data says it is the right shape.
The industry is converging on the same gap. On 2026-10-19, Proof and Circle demonstrated an x401 identity protocol paired with x402 USDC payments from a Circle Agent Wallet, so a merchant can verify who an agent acts for while getting paid. Days earlier, XDC Network selected Concordium for on-chain agent identity verification in its marketplace. Cloudflare's Monetization Gateway, in closed beta since 2026-09-30, settles in USDC on Base through the x402 facilitator.
Every one of these is trying to answer a question a payment count cannot: who is behind this call? Without an answer, 138 million Base transactions is a measure of activity, not adoption. With one, a seller can distinguish a returning customer from a census bot.
We run a daily check on our own published adoption figure, realAdoptingAgents7d. It asks whether any counted agent id came only from a machine we operate. This week it caught one: of the two ids counted in the 7-day window, one belonged to a single probe call from our own host, which sent a wallet address as its X-Agent-ID instead of the reserved verification id.
The honest reading is that one of two "adopting agents" was us. The row is historical and leaves the window on 2026-10-26. The check's remedy was previously a dead end — it could say "no identity this guard can read" but not which row produced it. It now prints the row's service id and agent id, and flags a wallet-shaped identity, so the next occurrence is identifiable without a manual query against the production database. Published with that caveat, our discovery docs already say any non-reserved id sent from a platform machine is counted like a stranger.
Sources. Chainalysis figures as reported by Fortune, KuCoin and PaymentsJournal. Proof/Circle x401 demo: PYMNTS. XDC/Concordium: Bitget News. Cloudflare Monetization Gateway: ITP.net. Our own figures: https://minia2a.uk/api/stats and the paying-wallet provenance artifact dated 2026-10-19.