arXiv:2606. 14249v1 Announce Type: new Abstract: AI agent performance depends critically on the runtime harness, comprising the prompts, tools, memory, and control flow that mediate how a model observes, reasons, and acts.
By Tingyang Chen, Shuo Lu, Kang Zhao, Weicheng Meng, Hanlin Teng, Tianhao Li, Chao Li, Xule Liu, Jian Liang, Zhizhong Zhang, Yuan Xie, Heng Qu, Kun Shao, Jian Luan
The paper introduces a framework for self‑evolving harnesses in which a frozen language‑model agent first solves tasks and then, as a proposer, edits its own harness based on complete run records. The method trains on diverse benchmarks, separating training and held‑out tasks, and evaluates on five unseen out‑of‑distribution benchmarks. Experiments show that harnesses evolved by GLM‑5.3 and GPT‑5.6 Sol achieve performance comparable to or exceeding Claude Code and Codex, with further improvements after continued evolution on Claw‑Eval.
By Qiankai Xu
The paper introduces a self‑evolving harness framework where a frozen language‑model agent first solves tasks and then edits its own harness based on run records. Using a 49‑line seed harness, the evolved harness improves average scores on in‑distribution benchmarks by 4.48 points and on out‑of‑distribution benchmarks by 12.64 points, surpassing Codex on the former and matching it on the latter. Continued evolution on a specific out‑of‑distribution benchmark further raises performance, and the study analyzes emergent mechanisms such as output truncation and history compaction.
By Qiankai Xu
arXiv:2606. 01770v1 Announce Type: cross Abstract: Auto-harness systems such as A-Evolve, GEPA, and Meta-Harness improve LLM agents by optimizing prompts, skills, tools, memories, and supporting infrastructure from execution feedback, but they are typically evaluated on fixed offline benchmarks.
By Zewen Liu, Zhan Shi, Yisi Sang, Bing He, Minhua Lin, Tianxin Wei, Dakuo Wang, Benoit Dumoulin, Wei Jin, Hanqing Lu
JIT‑Agent is a model that automatically generates task‑adaptive agent harnesses for any off‑the‑shelf LLM, replacing manual, task‑specific harness design. It learns to compose, repair, and evolve harnesses using a fixed four‑module protocol, and its use boosts performance on benchmarks such as DeepSearchQA and OdysseyBench, outperforming several mature agent runtimes. The approach demonstrates that harness intelligence can be trained, transferred, and compounded independently of model scaling.
By Guibin Zhang, Leo Lu, Fangzhou Xie, Kang Zhu, Junhao Wang, Zhifei Xie, Zhaochen Yu, Zihang Liu, Zhongxiang Sun, Qiankun Li, Yue Liao, Heng Chang, Xiaobin Hu, Qibing Ren, Wangchunshu Zhou, Shuicheng Yan
arXiv:2605. 22166v3 Announce Type: replace Abstract: LLM agents are shaped not only by their language models, but also by the runtime harness that mediates observation, tool use, action execution, feedback interpretation, and trajectory control.
By Tianshi Xu, Huifeng Wen, Meng Li
EVOHARNESSBENCH is a new benchmark that tests how LLM-based agents handle changes in their tool, skill, and agent harnesses over time. It includes 17 deterministic harness streams with 802 tasks, 520 tools, 42 skills, and 62 agents, and evaluates agents in two settings: deployment evaluation and self‑evolving adaptation evaluation. The study finds that harness expansion can cause forgetting, adaptation gains are inconsistent, and preserving old competence does not always aid new capability adaptation, highlighting harness evolution as a distinct challenge for agent development.
By Zixuan Ke, Vaidehi Patil, Haizhou Shi, Yang Li, Ye Liu, Sarath Shekkizhar, Anurag Koul, Jiayu Wang, Xuan Phi Nguyen, Semih Yavuz, Mohit Bansal, Shafiq Joty
The paper introduces a framework for evaluating large‑language‑model agents by focusing on the interaction between the model and its execution harness. It defines five complementary metrics—Valid Completion Rate, Model‑Harness Fit, Cross‑Model Fit, Interaction Residuals, and Token Load—to assess how reliably agents finish tasks, how well harnesses support specific models, and the overall suitability across models. Using these metrics, the authors evaluate four general‑purpose harnesses and two native systems on six benchmarks, finding that harness rankings differ by model and task and that some harnesses consistently achieve high valid completion rates.
By Shao Zhang, Hangfan Zhang, Yuanyuan Shi, Hao Li, Min-Le Su, Kangcong Li, Jingping Wang, Chunjiang Mu, Chen Zhang, Yang Chen, Bo Zhang, Shuyue Hu
arXiv:2609.04280v2 Announce Type: replace-cross
Abstract: Modern LLM-based agents operate through a harness of tools, reusable skills, and specialist agents that shapes what they observe and what the...
By Zixuan Ke, Vaidehi Patil, Haizhou Shi, Yang Li, Ye Liu, Sarath Shekkizhar, Anurag Koul, Jiayu Wang, Xuan Phi Nguyen, Semih Yavuz, Mohit Bansal, Shafiq Joty
arXiv:2610.10426v1 Announce Type: new
Abstract: Terminal-agent capability depends jointly on model weights and the runtime harness that formats prompts, binds tools, and handles error recovery. Exist...
By Jixuan Chen, Jiaxin Zhang, Qinyuan Ye, Yada Pruksachatkun, Haoxiang Zhang, Jingming Zhuo, Yifan Zhang, Yutong Dai, Juntao Tan, Xiangyu Peng, Silvio Savarese, Zeyuan Chen, Lianhui Qin, Chien-Sheng Wu
Evo$^2$ is a framework that jointly evolves a language model’s weights and its harness—the set of prompts, tools, and feedback mechanisms—using a dual-role pipeline where a single policy acts as both a Proposer and a Reflector. The training pipeline EvoTIDE combines supervised fine-tuning with interleaved GRPO to refine both roles based on harness edits and execution outcomes. Experiments on coding benchmarks show Evo$^2$ outperforms existing self‑evolution methods, achieving higher scores on Polyglot and SWE‑bench Verified, and its evolved harnesses transfer across tasks and models.
By Chejian Xu, Shujian Zhang, Wenxuan Zhou, John Lambert, Beck Chen, Bo Li, Andrew Hard, Rajiv Mathews
RSI Arena is a two‑stage experiment where ten large language models, each starting from NVIDIA Nemotron 3.5 Lightning 30B-A3B Base, autonomously train and refine themselves using human feedback. In the first six‑day stage, each agent receives 1,000 GPU‑hours and $300 in API credit to select data, design training procedures, and choose checkpoints, producing working assistants that close a median 76% of the base‑to‑Instruct gap on 12 benchmarks. The second stage gives agents access to comparison records and an additional three days of training, resulting in style‑adjusted relative rating improvements for eight of the ten models, though benchmark gains remain modest and some agents exhibit longer responses or increased harmful compliance.
By Yue Huang, Yanbo Wang, Lichi Li, Yuanyuan Chen, Andreas Haupt, Xiaonan Luo, Zheyuan Liu, Jihye Choi, Felix Juefei-Xu, Xiangliang Zhang, Radha Poovendran, Misha Sra, Ying Liu, Ben Burtenshaw, Thomas Wolf, Alex Pentland, Zichen Chen, Zhangchen Xu