The Agents area indexes 289 providers — AI agents, agent frameworks, and the runtimes that give them autonomous reasoning, tool use, and multi-step task execution. It is the newest of the 77 areas and, on the evidence, the most consistently high-scoring young cohort in the catalog.
The standouts
| Provider | APIs | Composite | Band |
|---|---|---|---|
| Vapi | 15 | 75.7 | exemplar |
| ChatGPT | 2 | 75.6 | exemplar |
| Anthropic | 24 | 75.0 | exemplar |
| GitHub Copilot | 12 | 75.0 | exemplar |
| Exa | 19 | 74.9 | exemplar |
| Dust | 14 | 72.0 | exemplar |
| Buildkite | 23 | 71.8 | exemplar |
| Landing.ai | 13 | 71.4 | exemplar |
| Daytona | 28 | 70.3 | exemplar |
| Parallel | 6 | 69.2 | exemplar |
| Microsoft Copilot | 7 | 66.8 | exemplar |
| E2B | 14 | 66.6 | exemplar |
| WorkOS | 41 | 66.3 | exemplar |
Thirteen exemplars, and the largest surface among them is 41 APIs.
Nobody here has a big API
That is the defining characteristic. Compare the top of almost any mature area — observability, security, commerce — and you find providers with 80, 130, 200 APIs. The Agents area’s leaders publish a median of around fourteen.
Two things explain it. First, these are young companies; there has been no time to accumulate surface. Second, and more interestingly, an agent-facing product does not want a large API. The whole design goal is a small set of composable, unambiguous operations that a model can select between correctly. Two hundred endpoints is a liability when your consumer has to pick one from a tool list.
ChatGPT scores 75.6 on 2 APIs. That is the shape of the category.
The area is really three products
Read the list again and it separates cleanly:
Models and assistants — ChatGPT, Anthropic, Microsoft Copilot, GitHub Copilot. The reasoning layer.
Agent runtimes and sandboxes — Daytona, E2B, Buildkite. Where the agent’s code actually executes, safely, with a filesystem and a network it cannot escape. This is the least-discussed and most operationally serious part of the stack.
Tools the agent calls — Exa (retrieval), Parallel (retrieval), Landing.ai (vision), Vapi (voice), WorkOS (identity and enterprise auth), Dust (orchestration over company data).
An actual deployed agent is one from each column. The area’s coherence comes from the fact that all three are being built by the same cohort at the same time, for each other.
WorkOS is the one to notice
WorkOS at 66.3 with 41 APIs is the largest surface in the area, and it is an identity and enterprise-auth company — not an AI company at all.
It is here because the unsolved problem in agent deployment is not reasoning. It is authorisation: which agent, acting for which user, may call which tool, with what scope, and how is that revoked. That is an identity problem with thirty years of prior art, and the companies who already solved it for humans are being pulled into this cohort to solve it for machines.
Takeaway
289 providers, thirteen exemplars, and not one large API among them. The Agents area is a young cohort scoring well precisely because small, unambiguous surfaces are the design requirement — and its biggest surface belongs to an identity company, which tells you where the hard part actually is.
Browse it at agents.apievangelist.com, and the full set at apis.io/areas/.