You let the smarter model explore the traces and figure out where the current harness' bottlenecks are for the current LLM. Then you can adjust prompts or tools to fix those.