An arXiv preprint reports adaptive validation-task selection that lowered AI-agent benchmark search costs while keeping performance near full search.