Google Research 开源了 Harness²，让 Agent 自己去改外面那层配置，跑一轮下来在 WorkBuddy-Bench 上最多涨了 11.5 分。
This is a dev post classified by AGK Intelligence as AI dev tools (a tool drop), kept by the Dev Radar because it carries real work, not commentary.
Google Research 开源了 Harness²，让 Agent 自己去改外面那层配置，跑一轮下来在 WorkBuddy-Bench 上最多涨了 11.5 分。 它把模型外围拆成 8 个能改的部件：系统提示词、护栏、记忆、工具说明、子 Agent、Skill、脚本和插件。 做法是先提出几种改法，在同一个任务上各跑一遍，再根据几次执行的差异拼出更好的配置。整个过程不需要标准答案。 每个任务跑完，最后采用的配置和有证据支撑的经验都会存下来，后面的任务接着用。 测试模型是 GLM-5 和 Gemini 3.5 Flash。JobBench 最多提升 4.5 分，LAB 部分通过率最多提升 11.8 个百分点。 上手前提醒一句：执行器默认不限制联网。评测前最好先锁网，免得 Agent 直接搜到评分标准。 🔗 https://github.com/google-research/harness2
Posted by lumxss (34.5k followers) 1 h ago · 4 likes · 336 views · view the original post on X. Kept by the Dev Radar as AI dev tools. Tools mentioned: harness2.
More dev work like this
- Anthropic, the famously pro copyright company — @ThePrimeagen
- Impressive work from @pranavctl on cloud coding agents. — @omarsar0
- SpaceXSI is testing a new sharing disclosure in Grok Bot. — @blankspeaker
- Say goodbye to usage caps. — @llama_index
- “AI is now improving the systems that run AI.” Thanks to @Forbes for covering Together… — @togethercompute
- Hy4 beat DeepSeek at remaking the GTA VI trailer 🎬 — @aimlapi
- Stop rebuilding your LinkedIn writing prompts from scratch. — @DanKornas
- The real test of AI coding is what happens when the next agent builds on it. — @wallstengine
Every post is read and classified by AGK Intelligence: what it is, which market it belongs to, and whether the link is a real tool. 31.3k posts from 5k X accounts over the last 21 days, 9.3k tools, 12 markets. New posts are collected several times an hour and the full ranking is recomputed every hour — last update 2026-11-04 18:57 UTC. Full method.