Harness-of-Harness：持久状态与证据驱动提升长程代码Agent
Harness-of-Harness beat repeated coding-agent runs by carrying code, QA evidence…
长程代码生成是 Agent 落地难点，这篇论文给出了具体的工程解法：用持久状态和证据驱动替代盲目重试，对做 Coding Agent 的同学有直接参考价值。
Harness-of-Harness beat repeated coding-agent runs by carrying code, QA evidence, and plans forward.
Harness-of-Harness（HoH）通过在代码、QA 证据和计划之间进行传递，击败了多次重复运行的编码智能体。
Says long-horizon coding is not just about giving an agent more time; it needs persistent project state, independent testing, and replanning from real failures.
文章指出，长周期编码不仅仅是给智能体更多时间；它需要持久的项目状态、独立的测试以及基于真实失败的重规划。
The problem: over a long project, coding agents can forget earlier decisions, repeat work, break working features, and miss unfinished requirements.
问题在于：在长期项目中，编码智能体可能会遗忘早期的决策、重复工作、破坏已正常工作的功能，并遗漏未完成的需求。
Harness-of-Harness, or HoH, fixes this by carrying the current software, test evidence, known problems, and an updated plan into the next coding run.
Harness-of-Harness（或称 HoH）通过将当前软件、测试证据、已知问题和更新后的计划带入下一次编码运行来解决这一问题。
It improved all 3 tested agent setups across 3 software benchmarks.
它在 3 个软件基准测试中提升了所有 3 种被测试的智能体配置的表现。
The clearest comparison used Codex + GPT-5.5 on GameCraft-Bench: after 3 passes, HoH scored 71.52, while simply continuing the same coding agent scored 58.24.
最清晰的对比使用了 GameCraft-Bench 上的 Codex + GPT-5.5：经过 3 轮迭代后，HoH 得分为 71.52，而简单地继续运行同一编码智能体的得分仅为 58.24。
The paper also ran HoH for 70 loops and produced a playable FPS from high-level requirements.
论文还让 HoH 运行了 70 个循环，并从高层需求中生成了一个可玩的 FPS（第一人称射击游戏）。
That longer test is only 1 game project and used extra tools and skills, so broader real-world generalization is still open.
该较长测试仅涉及 1 个游戏项目，且使用了额外的工具和技能，因此更广泛的现实世界泛化能力仍有待验证。
So for long-running coding agents, invest in persistent project state, independent QA, and evidence-driven replanning—not just more calls.
因此，对于长期运行的编码智能体，应投资于持久的项目状态、独立的 QA（质量保证）以及基于证据的重规划，而不仅仅是增加调用次数。
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