Industry

模型变强后,agent harness 哪些环节可以砍掉

Why we picked it

几位 agent 圈的实践者在 Code 现场聊了个真问题:模型越强,你写的编排代码有多少可以删掉。Pi 现在就在做极简 harness 这条路,跟重框架派是两种打法。天天搭 agent 工作流的人看完会想回去审视自己的代码。

aiDotEngineer 在 Code 活动后发起讨论:随着模型能力提升,围绕 agent 的 harness(外围脚手架)里哪些部分可以移除。Vela Gao 在现场与 mitsuhiko 交流了这个话题,核心观点是当模型足够强时,harness 有机会保持足够简单。她同时提到 pidotdev(Pi)目前的产品思路正是朝着简化 harness 的方向在做。

Body

这场讨论的实质是 agent 工程的路线之争:一套思路是靠复杂的编排、工具链和校验层弥补模型短板,另一套思路是让模型能力本身消化这些工作,把 harness 精简到接近裸模型加少量接口。Pi 走的是后一条路,mitsuhiko 的开源实践(如 minpilot 一类极简 agent 实现)也持类似立场。

对开发者而言,这个问题决定了自己写的编排代码有多少是临时拐杖:如果模型每次升级都会淘汰一层 harness,过度工程化的 agent 框架反而会成为维护负担。这也是 Claude Code、Pi 等产品在设计取舍上反复被讨论的原因。

Original · AI Engineer

One of the conversations worth continuing after Code: as models improve, which parts of the harness can we remove? @vela_gao brought that question into the hallway. Vela @vela_gao So glad to meet @mitsuhiko in person and talked with him about the imagination of agent harness in the future when models are capable enough. It could still be simple enough, as what @pidotdev is doing now! @aiDotEngineer 🔗 View Quoted Tweet📊 2 ⚡

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Industry" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include agent harness, mitsuhiko, Pi, Claude Code.

Source attribution: curated from AI Engineer (@aiDotEngineer).