EnvOpt 论文：优化智能体运行环境，Haiku 4.5 通过率提升 14.34%
做智能体的朋友看看这篇：不改模型，只改智能体工作的环境，Haiku 4.5 通过率涨 14.34%，成本还能降七成。
一篇关于智能体环境优化的论文提出 EnvOpt 方法，目标对象是运行在 Slack、文档、工单等共享工作区中的智能体。它用一个编码智能体在无评分器、无参考答案的训练任务上运行目标智能体，编辑环境并根据验证集决定停止。在 AutomationBench 上，该方法让 Claude Haiku 4.5 的通过率提升 14.34%，成本降低 72.7%。同一环境迁移到 Opus 5 后成本下降 68.1%；叠加 Meta-Harness 的优化 harness，成本再降 83.4%，准确率增加 11.17%。
Brilliant paper on optimizing agent environments.
This is an extremely important skill if you work with agents today.
You can make a lot of improvements by optimizing the environments agents operate in.
Many agents work inside a shared workspace, such as a company's Slack or a set of documents and tickets. EnvOpt changes what that workspace holds and how it is presented to the agent.
A coding agent runs the target agent on training tasks without graders or reference answers, edits the environment, and stops based on a validation set.
Haiku 4.5's pass rate also rises by 14.34%. The same environment transfers to Opus 5 with 68.1% lower cost. On top of Meta-Harness's optimized harness, it lowers cost by another 83.4% and adds 11.17% accuracy.
On AutomationBench, it cut Claude Haiku 4.5's cost by 72.7%.
https://t.co/wclYaAaK4H
Chat with Paper: https://t.co/eObEZAKyIu
