x402 is no longer a small protocol experiment.
On August 28, 2026, x402’s official dashboard showed 75.41M transactions, $24.24M in volume, 94.06K buyers and 22K sellers over 30 days.
Those “buyers” are protocol addresses, not a verified census of autonomous AI agents. The public data cannot tell us how many were agents, developers, scripts or test wallets. The shift is still important: Coinbase, Cloudflare and Circle are building systems in which software can discover a service, authorize a bounded payment and continue without a checkout flow designed for a person.
AI agents are becoming a new kind of paying customer. The useful question for a builder is not whether every wallet is autonomous today. It is what software will repeatedly pay for when this behavior becomes normal.
Then we looked at how that activity was distributed across the services agents can discover.
The internet now has a catalog of things software can pay for by itself.
In our August 28, 2026 snapshot, Coinbase's x402 discovery API listed 14,812 paid resources across 1,614 hosts. Search, market data, verification, model access and dozens of narrower services were available behind a machine-readable price.
That sounds like a market.
Then we looked at the middle of it.
The median active endpoint reported one call in 30 days. It also reported one unique payer.
The interesting question is no longer whether an agent can pay an API. It can.
Can an agent tell what is worth buying?
That is the gap we followed.
What x402 changes
The normal web assumes a person can create an account, accept terms, enter a card and manage a subscription. An autonomous process cannot reliably do that for every service it encounters.
x402 revives HTTP's unused 402 Payment Required status as a payment handshake. A server tells a client what a resource costs; the client signs a stablecoin payment; a facilitator verifies and settles it. Coinbase's Bazaar adds discovery, so a machine can search for payable services before it knows their URLs.
Circle says its Agent Stack gives agents wallets, payments and access to more than 900 paid services, with USDC accounting for roughly 99.8% of x402 payment value at the time of its announcement.
The rails are real. But rails and markets are not the same thing.
What we measured
We queried Coinbase CDP's public x402 discovery endpoint on August 28, 2026 and deduplicated the returned resource URLs and hosts. The catalog fields include calls and unique payers over a trailing 30-day period.
| Catalog measure | Snapshot |
|---|---|
| Listed resources | 14,812 |
| Unique resource URLs | 14,809 |
| Unique hosts | 1,614 |
| Reported calls, 30 days | 291,153 |
| Endpoints reporting activity | 14,765 |
| Median calls among active endpoints | 1 |
| Median unique payers | 1 |
| Active endpoints with one payer or fewer | 74.1% |
| Active endpoints with 10+ payers | 232 · 1.6% |
| Share of calls on the top five hosts | 47.1% |
| Median listed price | $0.02 USDC |
These are catalog-reported activity fields, not audited revenue. A payer may be a developer, script, test wallet or autonomous agent. Multiple calls may come from one integration. We cannot turn them into “customers” without inventing facts the dataset does not contain.
The protocol-wide x402 dashboard and Coinbase’s discovery catalog measure different surfaces. The first counts settlement activity across x402. The second describes discoverable resources and their provider-reported usage. We use the dashboard to establish scale and the catalog to examine how demand is distributed; we do not treat their totals as interchangeable.
That caveat makes the result more useful, not less.
A lot of supply, very little proof of repeat demand
The catalog is enormous because publishing supply is becoming cheap. Discovery, payment and settlement no longer need to be rebuilt by each provider.
Demand is much more uneven.
One Chainlink workflow reported 36,333 calls from one payer. An X search service reported 28,940 calls from 27 payers. An Exa-powered search endpoint reported 20,571 calls from 244 payers. Those are three different shapes of activity hiding under the same headline number.
A high call count can mean product demand. It can also mean one integration polling frequently. Unique payers improve the picture, but still do not tell us retention, revenue quality or whether the buyer was autonomous.
The missing primitive is not payment anymore. It is confidence.
Before an agent spends even twenty cents, it needs answers a human normally supplies:
- Does this service actually return the data I need?
- Is the price competitive for this exact task?
- Has more than one buyer used it?
- What happens when it is wrong or unavailable?
- Can I send sensitive input to this provider?
- Is the purchase inside my mandate?
Search results are not enough. A buyer needs a policy.
What can builders sell AI agents?
Not “AI” in the abstract. Sell a bounded result the agent can evaluate, purchase and use inside a larger task. Four product shapes look worth testing.
1 — Verified task-specific services
An agent does not need another general-purpose chatbot. It needs a result: verify this token, return a cited company brief, find flight availability, validate an address or monitor a market condition. The product is a structured answer with evidence the buyer can inspect—and a price attached to one completed request.
2 — Trust and quality checks
Calls measure usage, not quality. Agents could pay for an independent check that combines uptime, response validation, payer diversity, refunds and repeat purchase without pretending any one metric proves trust. The product answers one task-specific question: should I trust this service for this purchase?
3 — Outcome-based routing
The cheapest endpoint is not always the cheapest successful result. A router could test providers against an acceptance policy, compare their outputs and charge for delivering an accepted answer rather than merely forwarding a request.
4 — Human escalation
Some purchases end in ambiguity. An agent could buy a bounded expert review with a clear scope, price, deadline and accountable reviewer when confidence falls below its mandate. The product is not unlimited human help. It is a precise escape hatch for decisions software should not make alone.
The picks and shovels behind the market
There is another opportunity, but its customer is the API provider rather than the agent. Most APIs were designed for subscriptions and human developers. Metering, receipts, machine-readable terms, per-request privacy controls and dispute handling are still product work. Seller tooling may be valuable infrastructure for every product above; it just should not be confused with something an agent buys to complete its task.
What we built
We could have stopped at the catalog. Instead we picked one of the four products: a procurement layer that helps an agent decide what deserves its budget.
We gave a simulated agent $1 and asked what it should buy.
Agent Buyer is a dry-run procurement layer. Choose a task, set a budget and require a minimum number of reported payers. It compares a representative service set, rejects options outside the mandate and explains the proposed purchase plan.
It does not connect a wallet or call an endpoint. That is intentional. Sending a payment is the solved-looking part. Choosing responsibly is the experiment.
The market view keeps the uncomfortable contrast visible: 14,812 resources, a median of one reported call, and only 1.6% of active endpoints showing ten or more unique payers.
There is money for builders—but verify the terms
Circle Developer Grants explicitly includes agentic economic activity among the areas it supports. Coinbase has also advertised Founders Fuel, with CDP and AWS credits for eligible teams. Program availability and terms can change; check the live pages before planning around them.
Funding can subsidize an experiment. It cannot manufacture demand.
What this is not
This is not proof that 14,812 useful businesses exist, or that agents independently made 291,153 commercial decisions.
It is not proof that x402 failed either. Infrastructure markets often look over-supplied before buyer behavior catches up.
It is a snapshot of a strange transition: software can now discover and pay software, but it still needs help deciding whom to trust.
That is the part we would test next.