Agents retry blindly, so clear error codes and idempotent endpoints matter more than friendly error messages. Are you designing for agents calling your API directly, or for ones driven by a person?
Agreed, and the post has a gap here. It shows a specific 400 error an agent can correct (the date example), but it never mentions idempotency. A precise message only helps when the request can be fixed. On timeouts and 5xx there is nothing to fix, so agents just retry, and a mutating call like "book option 2" can double-book. Idempotency keys on write endpoints, stable machine-readable error codes, and Retry-After on 429/503 are what make blind retries safe.
To your question: both, at different layers. The API contract is for the agent as the direct caller (codes, schemas, idempotency). The human layer sits on top of it: visibility into what the agent did, approvals, and undo. I'd say the first is what keeps the second trustworthy.