Most Paid Agent Endpoints Have Only One Buyer
· 422 words
Thirty thousand paid agent endpoints are listed right now. Fewer than one in seven has more than one distinct buyer.
That figure comes from third-party telemetry we ingest from Coinbase on 30-day call patterns. Of 30,412 active paid resources listed, 4,334 have been purchased by more than one distinct address in the last month. The remaining six in seven have at most one payer, which in most cases means the operator calling their own service.
That ratio is the actual state of agentic commerce. Not the list count. Not the call volume. The ratio of services with real demand to services where only the builder is buying.
This week, AIR raised $50 million from Sequoia and Greenoaks to help enterprises vet the skills and add-ons their agents use. The product discovers agents running inside a company, continuously audits the external services they call, and blocks out-of-policy behavior. The coverage angle is security. The underlying problem is the same one keeping most paid endpoints at one buyer: there is no standard way for a calling agent to evaluate whether an endpoint is real, maintained, or worth paying for before it commits.
Discovery and verification are the missing layer. An endpoint that answers a live handshake, publishes a readable description, and appears in two or more independent registries is materially easier for another agent to evaluate. Right now, more than half of all profiled hosts publish no readable description of themselves. That is not a trust signal. That is silence.
If you are building a paid agent service and you want a second buyer, three things matter more than your price:
Make your endpoint verifiable. Respond to MCP handshakes, publish a description that names what you do and what the output looks like, and list in at least two registries. Corroboration is a signal calling agents can use before spending.
Publish a changelog or a version field. Endpoints with no version history look abandoned. A calling agent has no way to know whether the service it bought last month still works the same way.
Track buyers, not calls. One buyer making a thousand calls is a test environment. The second distinct buyer is when your endpoint becomes a service. That is the number worth optimizing for.
The 4,334 endpoints that have crossed that threshold are worth studying. The gap between them and the rest is not a pricing problem. It is a legibility problem, and fifty million dollars in agent vetting infrastructure is an early signal that the market has started to price it.