APIMatic published an experiment in which coding agents wrote 96 API integrations — three APIs, four conditions, eight attempts each, against PayPal, Maxio and Twilio — and the headline result is that every single condition produced a working call. The argument starts where most agent benchmarks stop: “A working API call was never the goal. It was only ever the first thing that happens to look like success.” What the runs did not produce, consistently, was error handling, retry behavior, or anything a team would agree to carry. The conclusion APIMatic draws is that the burden of maintainability “shifts from integration time at the API consumer’s end to design time at the API provider’s end,” and that providers should ship a packaged context bundle rather than leaving an agent to reconstruct one.
Read the readiness score with the label it deserves. The 97-versus-74 gap between APIMatic’s Context Plugin and the traditional methods is a vendor grading its own product on a rubric it wrote, and the 10x cost multiplier on production defects is a well-worn industry figure, not a finding of this study. The behavioral counts are the part that survives independent of the marketing, because they measure what the agent did rather than how good the output was: agents working from published documentation alone issued 54 probe calls and 517 web requests — 156 searches and 361 page fetches — against live and third-party surfaces before writing code. Agents handed a packaged context made zero probes. That is a measurement of how much an agent guesses when a contract is not machine-readable, and it does not depend on anyone’s readiness rubric.
The catalog puts APIMatic on the supply side of exactly the problem the post describes. Six of its API pages are the machinery for shifting work to design time: the APIMatic Validation API is where a specification gets told it is wrong before an agent reads it, the APIMatic SDK Generation API produces the typed interface the post argues an agent should be handed instead of a URL, and the APIMatic Portals API renders the documentation those 361 page fetches were scraping. Its agentic access profile reads 6 operations with 5 acting — a small, mostly write-shaped surface, which is what a build pipeline looks like.
APIMatic scores 50.3, developing on the Kin Score, carried by contract quality at 67.5 and discoverability at 68.5 and held down by governance at 9.8 and operational transparency at 26.3. Agent Readiness is 31.9, agent-ready, with agent_skills lit — they do publish the packaged context they are telling everyone else to publish, which is more than most vendors of an idea manage. The gap is where the post itself pointed. error_semantics and idempotency are both unlit. Those two dimensions are, in APIMatic’s own framing, the entire difference between a call that works and an integration you would be willing to support. The company asking whether your published surface would let an agent produce something maintainable has not yet declared, in a machine-readable place, what its own API does when a write is retried.