Agent Harnesses Reportedly Drive Large ARC-AGI Score Gains
An InfoQ 中文 AI&大模型 account of YC Paper Club's Harness Night says the same model weights can score from 30% to 95% or even 100% on ARC-AGI when the harness changes. Seth Karten described a harness as an expressive layer spanning model weights, active context with a REPL, and filesystem-like scheduling state, with compression, CRUD operations, and refinement connecting the layers. Other talks addressed hardware limits and organizational issues involving permissions, trust, and collaboration.
The reported results suggest that agent architecture, state management, and tool access can materially change benchmark performance even when the underlying model weights remain unchanged.
“Harness 就是包装纸，就是脚手架，就是提示工程”。这种论调在 Reddit 上每隔几天就会冒出来一次，甚至有人说“上下文工程根本不算研究问题”。但事实是，同一个权重文件，换一套 Harness，ARC-AGI 的成绩可以从 30% 直接飙到 95% 甚至 100%。
在 YC Paper Club 的 Harness Night 专场活动中，主持人VC Visiting Partner Francois Chaubard与几位论文作者，Prime Agent 作者 Seth Karten、OpenJarvis 作者 Jon Saad-Falcon、以及 QM 团队的 Josh France和 Regan Bell一起，围绕“为什么 Harness 比模型本身更重要”展开了三场演讲。连起来看是一条线：Seth 讲的是 Harness 该有什么样的表达力，Jon 讲的是这套东西能跑在哪种硬件边界内，Josh 和 Regan 讲的是把它放进真实组织后，权限、信任和协作会先出问题。本文基于该视频整理，经InfoQ编辑。
#YC Paper Club#Prime Agent#OpenJarvis#QM#ARC-AGI#REPL#Claude Opus#Gemini Flash#GPT-5.6 Sol#GPT-5.6 Terra#Anthropic#Nvidia