Why the console in the hero of this site skips RAG entirely, and when a system prompt is genuinely enough.

Notes on splitting 'did the tool call succeed' from 'was the answer actually good' in LangGraph eval workflows.