An zero-training system that enhances agent performance on medical tasks. It automatically evaluate LLM against golden references, identify weaknesses; upgrade prompts+RAG to avoid weakness patterns, and production API.