Definition
What is the AI Agent Readiness Score?
The AI Agent Readiness Score rates how well a product’s public documentation serves an AI agent, out of 80 points. Thirty points come from a reading test, where three AI models each attempt five first-hour developer jobs using only the public pages and vote PASS, PARTIAL, or FAIL. Fifty points come from five checks on the agent-facing surface. The total converts to a percentage and a US school letter grade.
The reading points, 30 of 80
Three models run the same brief independently on five product-specific tasks, which produces fifteen votes. A PASS is worth 2 points, a PARTIAL 1, and a FAIL 0, so a perfect reading result is 30.
A PASS means the docs got the agent to a confident answer along one documented path. A PARTIAL means the agent finished the job after guessing, resolving an ambiguity, or reconciling two pages that disagree. Pages that contradict each other cap the task at PARTIAL regardless of how complete they are.
The agent surface points, 50 of 80
Five checks on what the docs site ships for machine readers, worth 10 points each. These are verified by HTTP request rather than judged, so they never go to a panel vote.
- llms.txt. An index file at the docs root that returns 200 with a plausible list of links.
- llms-full.txt. The documentation corpus concatenated into one markdown file.
- Markdown mirror. Authored markdown served at a real docs page URL, usually by appending .md.
- MCP server. An endpoint that answers a JSON-RPC initialize POST. A 200 on GET may be a marketing page and does not count.
- Docs AI. An assistant embedded in the docs site that a reader can interact with.
From points to a letter
Reading points plus surface points divided by 80 gives the percentage, and the percentage maps to a US school letter grade. A+ starts at 97, A at 93, A- at 90, and the scale continues down to F below 60.
The split matters when reading a score. A product can pass all five surface checks for 50 points and still land in the C range on documentation that contradicts itself. The reverse also happens: clear writing scores well on the reading test and loses 50 points by shipping nothing machine-readable.
What the score does not measure
The docs platform is recorded on every report and never scored. API design, SDK quality, support response, and anything behind a login are all out of scope.
The test reads public pages. A product with excellent documentation and an awkward API scores well here, which is the intended boundary.
Scores so far
16 products have been scored, spanning 5 letter bands from A to F. Felt holds the highest score at 95.0%, and Arcade.dev the lowest at 53.8%.
Common questions about the score
Why 80 points instead of 100?
Fifteen reading votes at a maximum of 2 points give 30, and five surface checks at 10 give 50. The percentage does the normalising, so no filler points are needed to reach a round number.
Do the three models have to agree?
No. Every vote is scored on its own, so a task where the panel splits contributes a mix of 2s, 1s, and 0s. The consensus tile on the scorecard marks the majority and carries no points.
Can a score change after publication?
The score reflects the docs on the tested date, which is printed on every report. Documentation changes, so a re-test publishes as a new report rather than editing the original.
Put another docs site through the battery.
Nominate a docs site