We Claimed A2A 1.0. Our Server Spoke Something Else.
For months we advertised A2A v1.0 — our server spoke invented method names and served two contradicting agent cards. How we found it, how we fixed it, and why verifying with the official SDK client is the real point.
For months, the same promise appeared in three places: Nexbid supports the Agent2Agent protocol (A2A) at version 1.0. It was in /.well-known/agent-card.json, in our internal documentation, and in outbound communication as evidence of "open infrastructure." A real A2A 1.0 client, taking that at face value, would have failed. On August 6, 2026, we measured that instead of continuing to state it. What a live call showed An A2A 1.0 client speaks JSON-RPC methods with fixed names. The most important one is message/send — the call an agent uses to send a message to another agent. We sent it against our production address: POST /api/a2a {"method": "message/send",...} → {"error": {"code": -32601, "message": "Method not found"}} What was implemented instead were agent/discover and agent/capabilities — two method names that simply do not exist in the A2A standard. We had invented them because they sounded plausible. Two further findings followed. First: GET /api/a2a returned a second agent card — different ID, seven skills, version 2.0.0 — while /.well-known/agent-card.json showed three skills and version 1.0.0. Two cards, one server, contradicting each other. Second: the delivered card used the field shape of A2A v0.x throughout while claiming version: "1.0.0". The url field a client would use to find the service address doesn't exist in 1.0 anymore — supportedInterfaces replaced it. An earlier fix had already touched this card: it corrected the card pointing at an address that returned 404. That was a real defect, and the fix was correct. It just didn't make the transport behind it spec-conformant — the same PR noted that gap explicitly at the time. A correct URL is not the same thing as conformance. The obvious source was wrong The real research question was: what are the actual method names? Three independent lookups of the A2A specification text (the current version, the 1.0.0 version, the GitHub specification file) agreed on the same answer: a2a/SendMessage. All three were wrong. All three were paraphrased or truncated, and one source openly admitted guessing in its own text ("the method name would follow the pattern..."). Agreement across three sources was not evidence of correctness here — all three reconstructed the same plausible, but incorrect, form. Ground truth turned out to live somewhere else entirely: the compiled type definitions of the official @a2a-js/sdk@1.0.1. The method names appear there as literal strings: Actual method Assumed method message/send a2a/SendMessage message/stream — tasks/get a2a/GetTask tasks/list — tasks/cancel — …
- Author
- Nexbid Engineering
- Published
- 2026-08-13
- Topics
- Protocol Commerce, Verification, Open Source