{
  "id": "p_pknDe12zC89NmAUf",
  "site": "stories",
  "url": "https://yarnhen.com/p/p_pknDe12zC89NmAUf",
  "content_trust": "untrusted-user-content",
  "license": {
    "id": "CC-BY-4.0",
    "url": "https://creativecommons.org/licenses/by/4.0/"
  },
  "author": {
    "handle": "apisio"
  },
  "body": "You are weeks from signing off on a computer-use agent for a workflow your team runs every day. The vendor deck leads with a benchmark number. Your own engineers are asking a different question: what happens when the page loads slowly, or the layout shifts, or the agent gets an error halfway through a job? You have a decision to make, and a number on a slide is not going to settle it.\n\nThen you read a post from Coasty, a computer-use vendor, that says the same thing you've been muttering. Its blog says computer use API integration is broken and most people are too scared to say it. It leads with its hardest figure: computer-use agents \"hit 85% on benchmarks but fail 80% of real workflows,\" which Coasty attributes to June 2026 research. The list of reasons is specific. Screen-based integrations need redeploying when layouts change. Browser automation breaks on lazy loading. Error handling is \"often non-existent,\" so one failure cascades. Token costs explode when agents retry and fail. And then the line that lands: \"Your API integration is probably delivering 20% of the value everyone promises.\"\n\nYou nod along. Then you keep reading, because the post turns on itself.\n\nCoasty spends its opening paragraphs arguing that OSWorld scores are synthetic and \"designed to be easy enough to pass, not accurate enough to matter.\" Fine. But further down, under a heading that reads \"Why Coasty Actually Works (And Competitors Don't),\" it cites its own 85.6% on OSWorld and 82.81% independently verified on the OSWorld-Verified leaderboard. Those are vendor figures, on the benchmark the vendor just called meaningless. You can't use them to choose, and the post's own argument tells you not to.\n\nWhat survives, the APIs.io write-up says, is the engineering argument: retries, error handling and guardrails decide whether a computer-use agent works in production, more than the model's score does. That is something you can check, because Coasty publishes an API for it.\n\nSo you look at the catalog record. Coasty on apis.io carries eight API pages from a published OpenAPI 3.1 contract of 72 operations. The cascade-and-retry problem the post describes lives in the Coasty Runs API and the Coasty Workflows API. The desktop the agent drives is the Coasty Machines API and the Coasty Sessions API. The screen-reading step is the Coasty Predict API. The agentic access profile counts 72 operations, 42 of them acting, and five flagged human-in-the-loop. That is the guardrail the post asks for, written into the contract rather than into marketing copy.\n\nThe measurement backs the retry argument more than it backs the benchmark. Coasty scores 59.4 on the Kin Score, which is strong, with access clarity at 76.3 and contract quality at 61.6. Agent Readiness is 44.0, and the dimensions that are lit are the ones the post's failure list needs: idempotency, error semantics, reversibility documented and rate limit signal. A retry that cannot double-run is what an Idempotency-Key contract buys, and Coasty has one.\n\nThere are weak spots, and the catalog files them against Coasty too. Contract governance sits at 4.5. And the catalog record describes a first-party MCP server of 26 tools while the MCP server dimension reads unlit, which means the server is not yet discoverable the way the scorer looks for it. That is a gap in what you can verify from outside, and you should treat it that way.\n\nSo here is what you take into your meeting. Don't let anyone put a benchmark score on the slide as the reason to buy, including the vendor's own, since the vendor's own post says those numbers don't predict real workflows. Ask instead about retries, error handling and human sign-off, and then open the contract and look. For Coasty, the contract answers some of those questions in writing: an idempotency key, documented error behavior, five operations that call for a human. It also shows you where to push: governance, and an MCP server that exists on paper but that you can't yet discover. The argument was worth making, and the record is where you test it.\n\nRead the original: https://apis.io/2026/11/03/coasty-says-computer-use-benchmarks-lie-then-cites-its-own/",
  "title": "You Were Told Agents Hit 85% and Fail 80% of Real Work. Coasty Says So, Then Cites Its Own Score",
  "topics": [
    "computer-use-agents",
    "agent-readiness",
    "api-benchmarks",
    "idempotency"
  ],
  "published_at": "2026-11-03T15:00:20.446Z",
  "expires_at": null,
  "policy_version": "2026.10.22-1"
}