Newest papers

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

Zhuoshi Pan, ZhiYuan Feng, Yukuo Cen, Honglin Lin, Junru Lu, Zhongxia Chen, Shuo Yang, H. Vicky Zhao, Jie Tang · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.26770 · cited by 0 works (OpenAlex) · open-access PDF

French

Le développement de Harness Code joue un rôle crucial dans la construction d'agents d'IA capables. Les approches récentes améliorent les harnesses grâce à l'amélioration récursive de soi (RSI), mais leurs gains de performance viennent souvent avec des coûts d'exécution plus élevés. Nous introduisons HarnessPro, une méthode sans étiquette pour évoluer les harnesses pour améliorer les performances des tâches tout en réduisant les coûts d'exécution. Dans ce cadre, les harnesses candidates sont évaluées à l'aide de tests générés par LLM sans accès à des solutions de vérité de fond ou à des tests officiels. Les agents spécialisés diagnostiquent des trajectoires d'exécution, planifient des mises à jour avec

Machine translation (Meta M2M100); the English original follows.

Abstract

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. In this setting, candidate harnesses are evaluated using LLM-generated tests without access to ground-truth solutions or official tests. Specialized agents diagnose execution trajectories, plan updates with different performance-cost priorities, and implement them as candidate harnesses. Rather than returning a single harness, HarnessPro uses normalized weighted squared-distance selection to provide cost-focused (lite), balanced (medium), and performance-focused (heavy) harnesses for different deployment budgets. Selected harnesses enter the next iteration, where execution feedback guides further updates. Our experiments cover Terminal-Bench 2, SWE-Bench Pro, and ProgramBench, using GPT-5.5 with Codex and GLM-5.2 with Claude Code. Across these benchmarks using GPT-5.5 with Codex, the heavy mode of HarnessPro achieves an average pass-rate gain of 2.0 percentage points and an average per-task cost reduction of 26.1% relative to RHO. This advantage also holds when using GLM-5.2 with Claude Code. Code is available at https://github.com/pzs19/HarnessPro

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Computer science, Task (project management), Selection (genetic algorithm), Software deployment, Plan (archaeology), Code (set theory)

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange