Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Google Research opensources recursive agent harness framework Harness²

AI summary

Google Research released Harness² under the Apache-2.0 license, a framework that recursively optimizes an agent’s surrounding harness without changing model weights. Its two layers combine task-level propose–probe–compose refinement with a persistent improvement harness that records accepted changes and evidence. Paper-reported results show maximum gains of 11.5 points on WorkBuddy-Bench, 4.5 on JobBench, and 11.8 percentage points in LAB partial pass rate, but the report says these figures were not independently reproduced and are not uniform across configurations.

In-depth insight

Harness² turns agent-harness tuning into an explicit, reviewable optimization process covering prompts, guardrails, memory, skills, tool guidance, scripts, subagents, and plugins, while its unrestricted network access and disabled Codex sandbox require careful isolation during evaluation.

引言

"A harness is the system around a model that enables it to act as an agent. Harnessing is the process of optimizing that system."

这是「每日一个开源项目」系列的第 258 篇。今天的项目是 Harness²(读作 Harness Squared)—— Google Research 开源的递归式 Agent Harness 优化框架,23,451 颗 Star,2,354 个 Fork,Apache-2.0 协议,2026 年 10 月 6 日创建。

过去一年大家越来越清楚:同一个模型,装在不同的"外壳"里,表现可以差很多。这个外壳就是 Harness —— 提示词、工具、记忆、技能、编排逻辑,等等。但外壳基本靠人调:试一版、跑几个任务、再改。Harness² 想回答的问题是:能不能让 Agent 在遇到一个设计 Harness 时没见过的新任务时,自己改进自己的外壳,并且不需要标准答案反馈? 它的做法是在模型权重完全不变的前提下,做两层递归。

你会学到什么

  • Harness 与 Harnessing 的区别,以及 Harness² 暴露的八个可编辑组件
  • 两层递归:任务级的「提议—探测—合成」(Propose–Probe–Compose)和 Harnessing 级的「改进 Harness」
  • 并行递归与顺序递归的差别,以及 --k 预算怎么用
  • 快速上手:不需要任何模型凭证的本地 demo,和跑一次 LAB 基准的完整命令
  • 使用前必须知道的两件事:API 成本,以及评测时的网络访问风险

#Harness²#Google Research#GLM-5#Gemini 3.5 Flash#OpenCode#Codex#Vertex AI#WorkBuddy-Bench#JobBench#LAB

Original · 8 min readRead the original
Group chat

No comments yet. Start the conversation.