Tag
提出 Population-Centred Evolution (PCE)，通过将整个 agent 种群的成功、失败与自修改历史蒸馏为“population block”来指导父代 harness 的改进，在 SWE-bench Verified 上达到 72.1%、Polyglot 上达到 96.7%，显著优于群体进化基线。
一篇微软研究院等机构的论文提出 Self-HPO，一种交替优化 harness 与策略模型的递归自我改进框架，将 Qwen3.5-35B-A3B 在三个编码基准上的平均解决率从 57.8% 提升到 64.8%，且联合增益超过各自增益之和。
论文提出 Self-HPO，一个递归自我改进框架，交替优化编码智能体的 harness 与策略模型，使 Qwen3.5-35B-A3B 在编码基准上的平均解决率从 57.8% 提升至 64.8%。
HarnessPro 提出一种无标签的递归自改进方法，通过轨迹诊断与加权 Pareto 选择生成 lite/medium/heavy 三档 harness，在 Terminal-Bench 2、SWE-Bench Pro 等基准上相比 RHO 平均提升 2.0 个百分点通过率并降低 26.1% 单任务成本。
Google open-sourced Harness², a plug-and-play framework that turns any agent harness into a self-improving system by recursively proposing, probing, and composing harness edits without touching model weights. Tested with Gemini 3.5 Flash and GLM-5, it outperforms fixed harnesses in a single round and carries lessons across tasks.
论文提出 AutoOnboarder，一个由智能"Onboarder"编码代理监督的自主系统，通过迭代调整基础模型、提示词和 harness 源代码来适配任务，在基因组测序预授权表单提交任务中以最高低 29 倍的成本匹配或超越 Gemini 3.5 Flash 基线表现。
Google 的 Harness² 论文研究智能体在开放世界环境中对自身 Harness 的自进化，采用双层递归机制：无需标准答案即可通过行为对比筛选有效修改，并将修改过程本身总结为经验以指导未来任务，在多个基准上显著提升 GLM-5 和 Gemini 3.5 Flash 的性能。
Microsoft researchers propose ActiveSaddler, a method that selects training tasks for agent harness auto-tuners based on unresolved failure patterns, boosting pass rates by 4.4 points on GAIA2 and 7.5 on Terminal-Bench 2.0 while cutting tuning costs sharply.
Devin's latest update makes the AI software engineering agent 30-70% cheaper depending on mode, achieved by integrating the latest models like SWE-2 and optimizing its harness through batched tool calls and improved prompt caching, while Devin Fusion now leads FrontierCode 1.1 Extended at $0.60 per task.
The paper proposes AutoRef, a method that automatically optimizes the executable harness of an image-generation agent using a coding agent that rewrites harness code while keeping models frozen. The resulting AutoRef-Harness improves open-weight FLUX.2[klein] 4B from 5.72 to 7.37 on held-out four-reference MultiBanana tasks, matching or exceeding proprietary models like NanoBanana Pro and GPT-Image-1.5, and transfers across generators, benchmarks, and reasoning models without re-optimization.
A developer built a custom benchmark for coding AI agents and found that a 35B-parameter model outperformed a 120B-parameter one when the harness was optimized, highlighting the importance of tailored evaluation over generic specifications.
EdgeGen is a synthetic task generation framework that creates database-grounded edge-case tasks to improve tool-calling agents through fine-tuning and harness optimization, demonstrating consistent performance improvements.
This paper introduces Regularized Recursive Self-Improvement (RRSI) for AI agent harnesses, which applies regularization to prevent overfitting during recursive evolution, demonstrating performance gains on multiple benchmarks.
This paper studies agent harness optimization to improve LLM tool agents without retraining, focusing on prompts and tool-boundary middleware, and introduces a protocol and the PRISM optimizer for measurable gains.
This paper introduces HarnessEvo to decompose LLM agent harnesses into separately-evolvable slots, revealing that optimization value is localized in specific components like reflection/control, and that uniform budget-splitting is sub-optimal, advocating for targeted budget concentration.
The paper proposes WHALE, an alternating optimization method for jointly training model weights and harness code in AI agents, achieving significant performance improvements across search QA, math reasoning, and chess puzzles.
The article introduces HarnessOpt-Bench, a benchmark for measuring how LLMs can improve other AI agents' harnesses, and presents findings from 5 frontier models, showing that model choice has a greater impact than harness choice.
AutoSaddler is a Microsoft tool that automatically optimizes LLM-agent harnesses by diagnosing execution traces and applying structured updates to prompts, tools, and middleware, demonstrating significant benchmark improvements.
AutoSaddler is an automatic harness optimization framework that improves LLM agent performance on long-horizon tasks by iteratively updating harnesses using failure signals, achieving substantial gains on benchmarks like GAIA2 and SWE-Bench.
Task-CoEvolve is a novel approach for efficient LLM agent harness optimization that adaptively selects validation tasks to reduce evaluation costs while maintaining performance, achieving an 80% reduction in evaluations on benchmarks.