openJiuwen：基于动态编排的编程Agent框架
Nice paper showing just how far you can push an agent harness.
Agent 工程实践干货，展示了如何通过动态运行时编排而非单纯堆算力来提升 Coding Agent 性能，SWE-bench 成绩亮眼，值得做 Agent 的同学参考其架构设计。
Nice paper showing just how far you can push an agent harness.
一篇出色的论文，展示了将智能体框架（agent harness）的能力推向多远的程度。
In most setups, the default coding agent harness is static. Capabilities get wired in at design time, and the run has no way to change how it is being executed.
在大多数设置中，默认的编码智能体框架是静态的。功能在设计时就被硬编码进去，运行时无法改变其执行方式。
openJiuwen is an open-source harness built to fix that.
openJiuwen 是一个开源框架，旨在解决这一问题。
It reaches 82.6% on SWE-bench Verified and 87.19% on Terminal-Bench 2.1, ahead of the strongest official leaderboard entries by 3.4 and 3.39 points.
它在 SWE-bench Verified 上达到 82.6%，在 Terminal-Bench 2.1 上达到 87.19%，领先最强官方排行榜条目分别达 3.4 和 3.39 分。
A bit on how openJiuwen works:
关于 openJiuwen 的工作原理：
Rail-based composition lets developers assemble single agents, delegated sub-agents and swarm flows over one shared execution substrate.
基于轨道（Rail-based）的组合允许开发者在一个共享的执行基座上组装单个智能体、委托的子智能体以及群体流程。
Runtime evidence from semantic diagnostics, execution outcomes and task progress then reshapes context, feedback and task control while the run is still going.
来自语义诊断、执行结果和任务进度的运行时证据随后会重塑上下文、反馈和任务控制，而这一切都发生在运行仍在进行时。
The model policy stays fixed throughout, so the gains are attributable to the harness rather than to the model underneath it.
模型策略在整个过程中保持不变，因此性能提升归因于框架本身，而非其底层的模型。
Paper: https://arxiv.org/abs/2608.27969
论文：https://arxiv.org/abs/2608.27969
Chat with Paper: https://academy.dair.ai/papers/openjiuwen-beyond-static-harnesses-for-long-horizon-coding-agents-2608.27969
与论文对话：https://academy.dair.ai/papers/openjiuwen-beyond-static-harnesses-for-long-horizon-coding-agents-2608.27969
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力