Philipp Schmid(@_philschmid)

Can Agents improve their own harness on unseen tasks? The @GoogleResearch team released Harness², a ...

8.5内容质量
Can Agents improve their own harness on unseen tasks? The @GoogleResearch team released Harness², a ...

TL;DR · AI 摘要

Google Research推出的Harness²框架通过自动化改进冻结模型的harness，在基准测试中显著提升性能。

核心要点

  • Harness²通过四阶段自动化流程实现模型harness优化，包含轨迹记录、编辑生成、合并验证和知识沉淀
  • Gemini 3.5 Flash在WorkBuddy-Bench基准测试中性能提升8.4个百分点
  • 技能模块占所有有效编辑的50%，移除技能会导致5.3个百分点的性能损失

结构提纲

按章节快速跳转。

  1. §引言

    介绍Google Research发布的Harness²框架及其核心目标

  2. ·框架机制

    详细说明Harness²的四阶段自动化优化流程

  3. ›实验验证

    展示Gemini 3.5 Flash在多个基准测试中的性能提升数据

  4. ·技术分析

    解析技能模块在编辑过程中的关键作用及影响

  5. ›知识沉淀

    说明Reflector agent如何构建可复用的组件库

思维导图

用一张图看清主题之间的关系。

查看大纲文本（无障碍 / 无 JS 友好）
  • Harness²框架
    • 工作流程
      • 轨迹记录
      • 编辑生成
      • 合并验证
      • 知识沉淀
    • 实验结果
      • WorkBuddy-Bench +8.4%
      • JobBench +7.5%
      • 技能模块贡献50%
    • 技术特性
      • 冻结模型优化
      • 自动化编辑
      • 组件库复用

金句 / Highlights

值得收藏与分享的关键句。

  • Harness²框架使Gemini 3.5 Flash在WorkBuddy-Bench基准测试中性能提升8.4个百分点

    — 结果部分

    ⬇︎ 下载 PNG𝕏 分享到 X
  • 技能模块驱动约50%的有效编辑，移除技能会导致5.3个百分点的性能损失

    — 实验结果

    ⬇︎ 下载 PNG𝕏 分享到 X
  • 四阶段自动化流程包含轨迹记录、编辑生成、合并验证和知识沉淀四个核心环节

    — 框架机制

    ⬇︎ 下载 PNG𝕏 分享到 X
#AI框架#模型优化#自动化工具#Google Research
打开原文

Philipp Schmid on X: "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 insp… / X

Philipp Schmid

@_philschmid

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. Paper:

arxiv.org/abs/2610.14224

Code:

github.com/google-researc…

11:58 AM · Oct 19, 2026

·

7,055

Views

32

12

148

84