Exo 论文：让智能体在工作中实时自我改写 harness 代码
一篇叫 Exo 的新论文，让智能体干活时自己改 harness 代码，在 AppWorld 上省了 35-61% 的模型成本，做智能体的可以看看思路。
论文提出 Exo，一种支持在任务执行过程中自我学习与适配的智能体运行时架构。传统 harness 优化通常离线进行，需要代表性数据集和多次昂贵运行才能完成改进，Exo 把这一循环搬进正常任务执行，让智能体边工作边编辑自己的 harness 代码。一层小型保护层负责保存执行历史并控制生效的策略版本，使每次改动都可审查。在 AppWorld 和 SREGym 基准上，Exo 相同起始策略下把留出集模型成本降低 35 到 61%，正确率相当；AppWorld 平均任务时间缩短 13%，SREGym 中位诊断时间缩短 35%。
If you want to learn how to build self-improving harnesses, check out this paper.
They present Exo, an agent runtime architecture for in-harness learning and adaptation during ongoing work.
Harness optimization usually runs offline, so it needs a representative dataset and many expensive runs before the agent improves.
Exo moves that loop into normal task execution.
The agent edits its own harness code while it works. A small protected layer keeps the execution history and controls which policy version is active, so every change can be inspected.
On AppWorld and SREGym, Exo cuts held-out model cost by 35 to 61% against the same starting policy without self-editing, with comparable correctness. Mean AppWorld task time drops 13%, and median SREGym diagnosis time drops 35%.
https://t.co/9PeaIA7poo
Chat with Paper: https://t.co/o01LJYjBzz
