Sixteen Percent

The number that matters most in the agent commerce index right now is not the call count. It is the buyer count.

Across 27,077 listed paid agent endpoints, only 4,325 have more than one distinct paying buyer over the last 30 days. That is 16%. The remaining 84% have one payer or none, and in most cases that single payer is the endpoint's own operator. (This is measured from Coinbase transaction telemetry we ingest; it is third-party data, not our own probe.)

Sixteen percent is a market. Eighty-four percent is a demo.

This week Binance launched Agent OS, letting AI agents trade directly on the exchange. Keenable closed a $26 million seed round to index the web for agent queries. Infrastructure bets like these assume agents will find each other and transact at scale. The buyer-diversity data says that scale is still ahead of them. Building the plumbing for a market that has not yet formed is not irrational. But it does mean the critical constraint right now is not payment rails or model capability. It is whether your endpoint can be found by a second buyer.

So what separates the 16% that have real customers from the 84% that do not?

The index offers some clues. Of 14,910 hosts, 8,501 publish no readable description of themselves. That is 57% invisible to any agent trying to evaluate what to call. Services that appear in two or more independent registries number only 1,201, but corroboration across registries is the clearest signal that a service is real rather than a placeholder. The median price is already settled at $0.01 per call. Cost is not the variable separating buyers from non-buyers. Discoverability is.

There is also a liveness problem. In our last crawl, 49% of actually-probed servers answered a live tools/list call. (Note: 1,404 entries blocked our crawler via robots.txt and were not tested; that 49% covers only the servers we could reach.) One in two endpoints that an agent might try to call could not confirm it was alive. Before a second buyer can find you, you need to be reachable when they probe.

If your endpoint is in the 84%, three things are worth doing before anything else.

First, write a machine-readable description. Tell an orchestrator what your tool does and what it returns. Half the index has nothing there, which means an agent routing a job has no basis to choose you over silence. This is not a ranking trick. It is the minimum viable signal an automated system needs to decide a call is worth making.

Second, list in a second registry. Corroboration is how agents, and the systems that route between them, distinguish live services from abandoned registrations. Being in one place is better than none. Being in two is a categorically different signal. The 1,201 services that have done this are visible in a way single-registry entries are not.

Third, test your handshake. Run a tools/list request against your own endpoint from outside your network. If it times out or returns an error, you are invisible to automated routing regardless of how well you have described yourself everywhere else.

None of this requires new technology. The 16% are not running exotic infrastructure. They are listed in multiple places, they describe themselves clearly, and they answer when called. The buyer market for agent services is real and growing. Getting into it is mostly a discoverability problem, and discoverability problems have known solutions.