LangChain: Improving agents is a data mining problem

LangChain mines agent traces to identify failures and fine-tune judge models cheaper than frontier LLMs. The approach uses evals to hill-climb performance and improve agent reliability.