代码智能体的递归自我改进
Recursive Self Improvement for Coding Agents | Cline Blog
编程智能体的递归自我改进 | Cline Blog
Written by
作者
Published on
发布于
July 24, 2026
2026年7月24日
Preface
前言
Recursive self-improvement is the idea that models can iterate and improve on themselves to unlock singularity. We have achieved a version of that at Cline.
Recursive self-improvement 的理念是模型可以迭代并改进自身，从而解锁 singularity。我们在 Cline 实现了该理念的一个版本。
In the past few days, the response to Kimi K3 has been incredible. Building on that momentum, we optimized its performance with Cline and achieved SOTA results on Terminal-Bench 2.1 using one-shot recursive self-improvement.
在过去几天里，Kimi K3 的反响令人难以置信。乘此势头，我们使用 Cline 优化了其性能，并通过单次递归自我改进在 Terminal-Bench 2.1 上取得了 SOTA 结果。
We got an 88.8% score with the run costing $49.8. That surpasses Fable 5 at less than a tenth of the price: Fable spent $552, and even GPT 5.6 Terra spent $400 on Tbench2.1.
我们获得了 88.8% 的 score，运行成本为 49.8 美元。这以不到十分之一的价格超越了 Fable 5：Fable 花费了 552 美元，即使是 GPT 5.6 Terra 在 Tbench2.1 上也花费了 400 美元。
Aside from the system prompt + a few nudges, this was entirely self-contained. A single prompt kicked off 17 hours of continuous coding agent run with GPT-5.6-Sol as the leader model. One billion tokens total: 400 million spent by the coding agent, 600 million by repeated eval runs. See the prompt used here.
除了 system prompt 加上一些 nudges 之外，这完全是自包含的。单个 prompt 启动了长达 17 小时的连续 coding agent 运行，由 GPT-5.6-Sol 担任 leader model。总共消耗 10 亿 tokens：coding agent 消耗了 4 亿，重复的 eval runs 消耗了 6 亿。查看此处使用的 prompt。
We could have done the normal hill climbing we always do. Instead, the model made a PR on its own, a first-of-its-kind hill climb that turned a good score into a SOTA score. Note: We tried this experiment with Fable 5 too but its AI safety filter kept downgrading the model to Opus-4.8, and we simply gave up.
我们本可以做我们一直做的常规爬山算法。相反，模型自己提交了一个 PR，这是一种首创的爬山方法，将良好的分数转化为 SOTA 分数。注意：我们也用 Fable 5 尝试了这个实验，但它的 AI 安全过滤器不断将模型降级为 Opus-4.8，我们干脆放弃了。
Cline’s hill climbing experiences
Cline的hill climbing经验
Hill climbing isn't new to us as we've been doing evals and ...