Tag
Introduces Recursive Harness Self-Improvement (RHI), a method that iteratively refines prompt-level harness specifications for AI agents using pairwise feedback, improving performance and reducing inference cost by up to 60% on diverse machine learning research tasks.