Preprint reports 80% fewer evaluations for LLM-agent harness optimization

An arXiv preprint presents Task-CoEvolve, a method for optimizing LLM-agent harnesses by changing which validation tasks are sampled as the harness evolves. The authors say experiments on online text classification and Terminal-Bench 2.1 matched the final performance of full-set search while reducing optimization evaluations by 80% versus that approach. The method samples tasks where candidate harnesses disagree and estimates full-set scores from those partial evaluations. The work is a preprint, and the authors say code will be released.