HarnessAdapt-RL trains coding agents by adapting their harness during RL
A new arXiv preprint by Lingxiao Tang, Wei Tao, Yiwen Guo and Lingfeng Bao introduces HarnessAdapt-RL, a reinforcement-learning method that trains a coding agent's policy while also adapting its harness. A frozen teacher generates task-specific mechanism-level skills for flawed repair strategies, while a separate teacher model evolves tools and hooks, keeping each change only if it improves the observed pass rate or observed trainable rate without lowering either. With a 9B base model, the authors report pass@1 gains over GRPO of 9.8, 12.6 and 6.4 percentage points on SWE-bench Verified, Multilingual and Pro, and say it outperforms GRPO with the human-designed OpenHands harness; the results are self-reported and not independently verified.
The authors conclude that training produces a model-harness pair, indicating that coding-agent performance can depend on co-evolving the harness alongside the policy rather than on the policy alone.
Computer Science > Software Engineering
arXiv:2610.28069 (cs)
[Submitted on 20 Oct 2026]
Title:Training the Whole Agent, Not Just the Model
Authors:Lingxiao Tang, Wei Tao, Yiwen Guo, Lingfeng Bao
View a PDF of the paper titled Training the Whole Agent, Not Just the Model, by Lingxiao Tang and 3 other authors