Google Research 发布 Harness²：让智能体在冻结模型外自行改进运行框架
Google 把 harness 本身交给 agent 改：Editor 出编辑、Composer 合并、Reflector 存经验，Gemini 3.5 Flash 三个基准全涨，做 agent 工程的可以看看代码。
Google Research 发布了 Harness² 框架，在模型参数冻结的情况下迭代改进围绕它的 harness（运行环境）。流程分四步：先用基础 harness 跑任务记录基线轨迹，Editor agent 起草 2 个候选编辑（SI、Skills、hooks 等）并运行，Composer agent 对比新轨迹与基线后合并有效编辑，Reflector agent 把经验存入库供后续任务调用。Gemini 3.5 Flash 在 WorkBuddy-Bench 上提升 8.4 分，JobBench 上提升 7.5 分，且在全部 3 个基准上都因自我编辑 harness 而提升。消融显示 Skills 和脚本贡献约 50% 的编辑，去掉 Skills 会损失 5.3 分。论文与代码已公开。
Can Agents improve their own harness on unseen tasks? The @GoogleResearch team released Harness², a framework that allows improve an harness around a frozen model. How it works: 1️⃣ Agent runs the base harness on each task to record a baseline trajectory 2️⃣ Editor agent inspects the trajectory, drafts 2 candidate edits (SI, Skills, hooks,….), and runs both. 3️⃣ Composer agent diffs the new trajectory against the baseline, merges edits that fixed failures (within a cap), and runs the final harness. 4️⃣ Reflector agent logs verified lessons into file and saves the winning components to a library for future tasks to pull on demand. Results: - 💡 Gemini 3.5 Flash +8.4 on WorkBuddy-Bench and +7.5 on JobBench. - 🛠️ Skills and scripts drive ~50% of edits; dropping skills costs 5.3 pts. - 📚 Gemini 3.5 Flash improves on all 3 benchmarks when editing its own harness. Pape arxiv.org/abs/2610.14224 GI0 Cod github.com/google-researc… M9s 💬 13 🔄 2 ❤️ 24 👀 1391 📊 15 ⚡
