You can optimize LLM agent behavior (harness code) much faster by only evaluating tasks where the agent struggles or disagrees with itself, rather than testing everything every iteration.
This paper introduces Task-CoEvolve, a method that makes LLM agent harness optimization faster by intelligently selecting which validation tasks to evaluate at each iteration. Instead of evaluating all tasks every time, it focuses on tasks where the agent's performance is uncertain, reducing evaluation costs by 80% while maintaining the same final performance.