Harness：模型越来越强，而装模型的"壳"成为性能瓶颈
七月，Cline 公布了一个实验：他们发现在 Cline 上使用 Kimi K3 做基准测试时，成绩和官方公布的差距很大。为此，Cline 让 GPT-5.6 仔细分析基准测试的整个过程，结果发现五处性能下降都是 Cline 自身的问题引起的。比如模型清理后台进程时，一条宽匹配的杀进程命令偶尔会把 Cline 自己杀掉——agent 杀死了自己的监考官。修改完后，成绩从 77.5% 涨到 88.8%，复跑三次平均也在 87% 以上。这已经摸到 Kimi 官方团队自己调的壳的水平线（官方报 88.3%）。而且每次跑分花的钱，从 79 美元降到 49.8 美元，省了 37%。
In July, Cline published an experiment: they found that running Kimi K3 benchmarks inside Cline yielded scores far below official numbers. They had GPT-5.6 analyze the entire benchmark process and discovered five performance regressions caused by Cline itself. For instance, when models cleaned up background processes, a broad-match kill command would occasionally kill Cline itself — the agent murdered its own proctor. After fixes, scores jumped from 77.5% to 88.8%, with three reruns averaging above 87% — approaching the level of Kimi's own team-tuned harness (officially reported at 88.3%). Cost per run dropped from $79 to $49.80, saving 37%.
这不是说 Cline 不够强，实际 Cline 很不错——我的第一个 AI 编程工具就是它，他们做这类工具比现在常用的几家都早。Cline 的这个实验，也间接证明了我之前那篇《我为什么从 Claude Code 换到了 Pi Agent》的观点：模型用的壳，也就是 Harness，可能会限制模型能力的发挥。
This isn't to say Cline is weak — actually Cline is excellent. It was my first AI coding tool, and they were building this category before most of the tools we use today even existed. Cline's experiment also indirectly validates my earlier article "Why I Switched from Claude Code to Pi Agent": the shell around the model — the Harness — may be limiting what the model can actually deliver.
先说清楚"壳"是什么
First, let's clarify what a "shell" is.
直白些说：壳（Harness）就是你和模型之间的那层中间层，负责把你的话解释给模型听、安排模型干活、处理模型的输出。很多人天天在用的 Claude Code、Codex、Pi Agent、Hermes、OpenClaw 这些自然语言编程工具，本质都是壳。
Plainly: the Harness is the middleware between you and the model. It interprets your words for the model, orchestrates tasks, and processes the model's output. The natural-language coding tools many use daily — Claude Code, Codex, Pi Agent, Hermes, OpenClaw — are all fundamentally harnesses.
之前，很多人磕提示词，研究怎么写才能让模型更高效地工作。而现在，模型能力越来越强，对提示词的理解能力也越来越强。此时，能否发挥模型自身的能力，靠的是如何有效地让它工作。就如同，之前你请实习生，得把做事的规则、方法都写清楚；现在你请的是高级专家，不用告诉他怎么做事，给他安排好办公室、助理、明确好职权责任，才决定了他能不能干好。
People used to obsess over prompt engineering — studying how to write prompts for more efficient model output. Now, models are increasingly capable and their prompt comprehension keeps improving. At this point, unleashing a model's ability depends on how effectively you enable it to work. It's like: before, you hired an intern and had to spell out every rule and method. Now you hire a senior expert — you don't tell them how to do their job. You give them an office, an assistant, clear authority and responsibilities. That determines whether they can deliver.
我的真实体验就是：同样的 Kimi K3、DeepSeek V4，在 Claude Code 和 Pi Agent 上跑，感受完全不一样。之前的文章里我提过 Databricks CTO 的评测报告，里面有数据支持：Opus 4.8 在 Pi Agent 里跑比在 Claude Code 里跑便宜 2.08 倍，任务完成质量完全一样。GPT-5.5 跑同一个任务，在 Codex 里烧了 123.5 万 token，在 Pi 里只烧了 66.5 万，少了一半。Databricks 报告的核心结论：外壳对成本的影响，比模型本身还大。
My real experience: the same Kimi K3 or DeepSeek V4 feels completely different in Claude Code versus Pi Agent. I mentioned the Databricks CTO's evaluation earlier — backed by data: Opus 4.8 in Pi Agent cost 2.08× less than in Claude Code, with identical task quality. GPT-5.5 on the same task burned 1.235M tokens in Codex, only 665K in Pi — nearly halved. The Databricks report's core conclusion: the harness impacts cost more than the model itself.
那怎么选壳？我自己看三条。
So how do you choose a harness? I look at three things.
第一，看"完成任务花多少钱"，不看"每 token 多少钱"。
First: look at "how much to complete a task," not "how much per token."
第二，看它默认给你塞了多少东西。有的壳，模型还没干活，先烧掉一万多 token 的"说明书钱"；有的壳只给一千，剩下的空间全留给任务。系统提示越大，你每次对话就越在为它买单。
Second: see how much it stuffs in by default. Some harnesses burn over 10,000 tokens in "instruction manual overhead" before the model even starts working. Others keep it to about a thousand, leaving the rest for the actual task. The larger the system prompt, the more you're paying for it every conversation.
第三，看自由度：能不能换模型、能不能改规则、开不开源。开源的壳出 bug 可以修，甚至可以让 AI 帮你改——Cline 那个实验，本质就是"AI 修壳"。闭源的壳，你只能等官方修，等不到就忍着。
Third: check degrees of freedom — can you swap models, modify rules, is it open source? With open-source harnesses, you can fix bugs yourself, even have AI fix them for you — Cline's experiment was essentially "AI repairing the harness." With closed-source harnesses, you wait for official fixes or live with the pain.
壳不是万能的
The harness isn't omnipotent.
Cline 自己也承认：POV-Ray 渲染、游戏移植那几个任务，壳怎么改都救不回来。壳只能捡回"被壳绊倒"的分数，模型不会的，壳变不出魔法。
Cline themselves admitted: for tasks like POV-Ray rendering and game porting, no harness tweaks could salvage them. A harness can only recover points "lost to harness friction." What the model genuinely can't do, no harness can conjure up.
而且壳在弱模型上也没那么大用。有篇 2023 年的论文做过类似实验：GPT-4 上有效，GPT-3.5 和另一个小模型上反而变差。基础模型不够强，壳再好也白搭。
And harnesses don't help much on weak models. A 2023 paper ran similar experiments: effective on GPT-4, but counterproductive on GPT-3.5 and another small model. If the base model isn't strong enough, the best harness is useless.
所以准确的说法是：模型决定上限，壳决定你摸到多高。不是"壳比模型重要"，而是"模型够强之后，壳决定了你能发挥出它的几成"。
So the accurate framing: the model sets the ceiling; the harness determines how close you get to it. It's not "harness matters more than model" — it's "once the model is strong enough, the harness decides what percentage of that strength you actually unlock."
旗舰模型的时代，大家都够强了——这时候，壳的差距，就是你的差距。
In the era of flagship models, everyone is strong enough — at this point, the harness gap is your gap.
所以，选模型之前，先看看你打算让它住在哪个壳里。
So before choosing a model, first decide which shell it's going to live in.