Stop running everything through LLMs: How I fixed my research agent