HarnessPro evolves agent harnesses for better performance at lower cost
Original title: HarnessPro: Recursive Harness Self-Improvement with Pareto Optimization for Performance-Cost Trade-offs
With GPT-5.5 and Codex, heavy mode gained 2.0 pass-rate points and cut per-task cost 26.1% versus RHO.
HarnessPro is a label-free method for evolving AI agent harnesses through recursive self-improvement while trading off task performance and execution cost. Candidate harnesses are scored with LLM-generated tests, without ground-truth solutions, and specialized agents diagnose trajectories and implement updates. It returns cost-focused, balanced, and performance-focused harnesses rather than one result. The authors report the advantage on Terminal-Bench 2, SWE-Bench Pro, and ProgramBench with both GPT-5.5 plus Codex and GLM-5.2 plus Claude Code.
Unconfirmed: Pass-rate and cost figures are reported by the authors and not independently verified.
- Heavy mode averaged a 2.0-point pass-rate gain versus RHO with GPT-5.5 and Codex.
- That setup also cut average per-task cost by 26.1% relative to RHO.
- Selection uses normalized weighted squared distance across lite, medium, and heavy modes.
- Selected harnesses continue into the next iteration using execution feedback.
- Code is available at the authors' GitHub repository.
- Products
- CodexClaude Code
From the original article
1 min read at huggingface.co
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 task performance while reducing execution cost.
By Zhuoshi Pan, ZhiYuan Feng, Yukuo Cen · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.