Last week I wrote that your agent isn’t failing because it hallucinates — it’s failing because of rate limits. The capacity-engineering toolkit in that post — concurrency caps, backoff with jitter, fallback models, caching — is real and it works. Deploy it and your agent stops dying.

Then a commenter (ANP2) pointed out the thing the post...