The A2A Standard

About this paper

No specification in this series has a wider gap between its conversation and its artifacts. The Agent2Agent Protocol is properly hosted at the Linux Foundation, properly versioned, specified as code, and it ships a conformance kit most standards here do not have. Its specification repository collected 25,221 GitHub stars in sixteen months — an OpenAPI-scale audience at a fraction of OpenAPI’s age.

And when every reachable host in the API Evangelist catalog was probed for the artifact that makes any of it work, 65 came back. Of those 65, ten publish a card a conformant client can actually read. The rest ship capabilities as an array, or skills as an object, or nothing at the canonical path at all — not through carelessness, but because publishing an Agent Card is a one-way action with no response, no validator in the pipeline, and no registry looking over anyone’s shoulder. This report grades all 65, names every deviation, and follows the strangest finding in the data: that the platforms with the best API practices are the ones getting the card wrong, because they moved first and nothing has told them since.

What's inside

1. Executive summary — 65 cards, ten readable
2. What A2A is, and what it is not
3. The origin — a vendor contribution, done properly
4. What the body actually publishes
5. Version reality — the field that cannot be compared
6. The coalition — and what this report does not measure
7. The adoption gap — 65 of 22,341
8. Conformance — a good kit, and no path from it to the population
9. The manifest is outliving the protocol
10. The regulatory layer
11. The scoring caveat — pointed at the instrument
12. What would have to be true
13. The investable thesis

What you get