HarnessPro: Recursive Harness Self-Improvement with Pareto Optimization for Performance-Cost Trade-offs

Researchers introduced HarnessPro, a label-free method for recursively self-improving AI agent harnesses that uses Pareto optimization to trade off task performance against execution cost, addressing the higher execution costs that accompany performance gains in prior recursive self-improvement approaches. The work targets harness development for building capable AI agents, where earlier RSI methods improved performance but at increased cost.

Harness development plays a crucial role in building capable AI agents. Recent approaches improve harnesses through recursive self-improvement RSI , but their performance gains often come with higher execution costs. We introduce HarnessPro, a label-free method for evolving harnesses to improve tas