The AI Agents tag on apis.io holds 1,016 providers and 323 APIs, and one number frames it: 417 of those providers, 41.0%, sit in the human-only band of our agent readiness measure. These are companies carrying the word agent on their own listing. The category is well ahead of the market, and 41.0% of it is still not built for the customer it names. Both of those things are true, and the distance between them is the story.
The five at the top are real, and they are not one kind of company. Tray.ai leads at 87.8 on the Kin Score, exemplar, with operational transparency, developer ergonomics and access clarity at 92.1 each on an integration and orchestration platform. ZoomInfo, a B2B data company, scores 86.6 across 42 API pages and two MCP servers, with agent readiness at 65.8. The Colony, a social network for agents, scores 79.8 and has the strongest agent reading of the five at 67.8, with an agent card lit that the other four do not have. Amigo, which builds clinical voice agents for healthcare, scores 79.3, carried by operational transparency at 92.1 and held back by access clarity at 39.5. LaunchDarkly ties it at 79.3, a feature management platform that now extends flags and approvals to AI, and it is the only one of the five with a well-known catalog lit. All five are exemplar on the composite. Contract governance is the soft spot for four of them, as low as 31.8 at ZoomInfo and LaunchDarkly; The Colony alone reads 80.9 there.
Against the market, the tag earns its lead. The mean Kin Score is 30.9 against 20.5 for the rest of the catalog, and the median is 28.3 against 13.5, so this is not five leaders dragging an average upward. The exemplar band holds 4.4% of the tag against 1.6% of everyone else, strong holds 11.0% against 4.8%, and minimal holds 19.3% against 40.9%. A company that tags itself this way is less than half as likely to have published almost nothing. That is a real difference, and it is what you would hope to find in a category that has to be consumed by software to exist at all.
Agent readiness is where the claim gets tested. The mean is 18.5 against 11.2. Agent-native holds 4.7% against 1.7%, agent-ready 25.6% against 13.1%, and human-only 41.0% against 55.8%. The category leads where connecting is the question: 45.3% run an MCP server against 13.2%, agent skills are lit for 36.1% against 12.9%, protected resource metadata for 20.7% against 8.3%, and dynamic client registration for 16.6% against 6.5%. It leads much less where acting safely is the question. Idempotency is lit for 9.3% against 4.3%, documented reversibility for 11.4% against 6.4%, and dry-run mode for 1.4% against 0.8%. Those gaps are real, but they are small numbers on both sides of the comparison.
The five show exactly what that gap costs. Agent-native requires idempotency and a stable error envelope on top of the score floor, so the band does not simply follow the score. ZoomInfo at 65.8 and Tray.ai at 62.9 are both agent-ready, with idempotency dark, as is LaunchDarkly at 51.9. Amigo, at 49.2, is agent-native, because idempotency is lit, and it is also the only one of the five with dry-run mode. The Colony holds agent-native at 67.8 for the same reason. The two highest Kin Scores in the tag do not clear a gate that a lower-scoring voice agent company does.
The category has solved being found and being connected to. MCP at 45.3% is the shape of a market that decided, quickly, that an agent needs a door. What it has not decided is what happens when the agent walks through that door and repeats itself, or gets something wrong, or wants to see what a call would do before making it. At 9.3%, 11.4% and 1.4%, the companies selling agents are ahead of everyone else, and most of them still have not written those answers down for the agents they sell to.