Your Agent Found the Right Schema. Then Ignored It.

The assumption nobody tests Most agent evaluations quietly assume something that never happens in production: that the model is handed exactly the right information, and nothing else. We changed that one assumption — and watched a model drop from 58% to 18% . Retrieval was not the problem. The correct information was sitting in the prompt. The model just didn't use it. TL;DR — Standard agent evaluations hand the model exactly the right schema. Real retrieval does not. Changing only that cost 40 points of accuracy, and neither a larger model nor better prompting recovered it. A $150 fine-tuning