HarnessPro self-improves agent harnesses with Pareto cost trade-offs

What the research found

HarnessPro is an arXiv paper (2610.26770) on label-free recursive self-improvement of agent harnesses that returns lite, medium and heavy variants; heavy mode reportedly gains 2.0 pp pass rate and cuts per-task cost 26.1 percent versus RHO with GPT-5.5 and Codex, with code on GitHub.

Claims and how well the evidence supports them

1. HarnessPro is a label-free method that evolves agent harnesses using LLM-generated tests 99%
2. Heavy mode achieves an average pass-rate gain of 2.0 percentage points over RHO with GPT-5.5 and Codex 98%
3. Heavy mode reduces average per-task cost by 26.1 percent relative to RHO 98%
4. Code is available at github.com/pzs19/HarnessPro 98%

Evidence

• huggingface.co https://huggingface.co/papers/2610.26770
• github.com https://github.com/pzs19/HarnessPro
• arxiv.org https://arxiv.org/abs/2610.26770

Actionability rated low.