AI时代软件架构：用验证循环与系统约束驱动AI编码
Thinking in Systems, Shipping in Loops
直接给出了AI辅助开发下的新工作流范式（人类定约束+AI写码+验证循环），并提炼出系统设计的三个关键属性，独立开发者和技术负责人可直接参考调整团队分工与工程规范。
“Writing code by hand is no longer an economically productive enterprise for the vast majority of programmers working at the vast majority of companies. That’s today. By the end of the year, it will be virtually all domains, virtually all programmers, virtually all companies.”
“对于绝大多数公司的绝大多数程序员而言，手工编写代码已不再是一项具有经济生产力的工作。这是当下的情况。到今年年底，这将成为几乎所有领域、几乎所有程序员、几乎所有公司的现实。”
David Heinemeier Hansson, CTO of 37signals · Rails World, Austin · 23 September 2026
David Heinemeier Hansson，37signals首席技术官 · Rails World，奥斯汀 · 2026年9月23日
Software engineering have evolved to systems architecture. That was always the destination.
软件工程已经演变为系统架构。这一直是我们追求的目标。
This is the new software engineering : designing the systems that let AI write code correctly at scale.
这就是新的软件工程：设计那些能让AI大规模正确编写代码的系统。
We see this in our portfolio companies. Artemis engineers merged 2 pull requests per day in January, 6 by May & 16 by August, 30,000 in eight months.1 Dan Shiebler, their CTO, describes the division of labor plainly : every line of the platform is written by agents, & the engineers design systems, set constraints & review outputs.2
我们在投资组合公司中看到了这一点。Artemis的工程师在1月份每天合并2个拉取请求（pull requests），5月份达到6个，8月份达到16个，八个月内共计30,000个。1 Dan Shiebler，他们的首席技术官，直白地描述了劳动分工：平台的每一行代码都由智能体（agents）编写，而工程师负责设计系统、设定约束并审查输出。2
Great design scales. Lauren Tan, an engineer on SpaceXAI’s Grok team, ships 2,000 pull requests a month to production, nearly one hundred a working day.3 Asked what makes that possible, she does not name a model but verification, loops that enable AI to validate its work.
优秀的设计能够规模化。SpaceXAI的Grok团队工程师Lauren Tan每月向生产环境提交2,000个拉取请求，平均每个工作日近一百个。3 当被问及是什么让这一切成为可能时，她没有提及某个模型，而是提到了验证循环（verification loops），这些循环使AI能够验证其工作成果。
Dartmouth professor Donella Meadows argued that the most powerful places to intervene in a system are the least intuitive, & that people reliably find them & then turn them the wrong way.4
达特茅斯学院的教授Donella Meadows认为，系统中干预效力最强的地方往往是最不直观的，而且人们往往会可靠地找到这些地方，然后朝错误的方向去操作它们。4
Three properties define great system design : resilience, self-organization & hierarchy.
三个特性定义了优秀的系统设计：韧性（resilience）、自组织（self-organization）和层级（hierarchy）。
Resilience is AI checking its own work, robustly, through tests, a reviewer agent & production observability rather than one gate. Self-organization is the loop learning, so a class of failure becomes a new skill the agent applies next time. Hierarchy is the layering of skills, tools & sandboxes into composable components that are reused once verified to work.
韧性是指AI通过测试、审查智能体和生产环境可观测性来稳健地检查自己的工作，而不是依赖单一的关卡。自组织是指学习循环，使得某一类故障转化为智能体下次应用的新技能。层级是指将技能、工具和沙箱分层组合成可复用的组件，一旦验证可用便进行复用。
This work isn’t easy. Systems architecture never has been. It was my goal in 2005 as a Java software engineer just as much as it is today as we build out our agentic systems.5
这项工作并不容易。系统架构从来都不容易。作为Java软件工程师，这是我2005年的目标；如今，当我们构建代理式系统（agentic systems）时，这依然是我的目标。5
“A new career as a professional maker of things. Yes, you will no longer be chiseling that code by hand, but you will be making amazing things. You will be steering intelligence.”6
“一种作为专业造物者的新职业。是的，你不再需要手工雕刻那些代码，但你将创造出令人惊叹的事物。你将驾驭智能。”6
- Artemis Security : 30,000 pull requests in eight months, an average of 16 merged per engineer per day in August 2026, up from 6 in May & 2 in January. ↩︎
- Dan Shiebler, co-founder & CTO, Artemis Security : “Every line of code in our platform is written by AI agents. Our engineers design systems, set constraints, and review outputs.” ↩︎
- The Complete Guide to pstack, Pt. 1, Lauren Tan, 1 September 2026 : “It’s allowed me to ship 2,000 PRs a month to production with high confidence.” Her verification-skill example is MIT licensed. The New Stack profiled the workflow in One engineer shipped 2,000 PRs a month to production, 19 September 2026. ↩︎
- Thinking in Systems : A Primer, Donella H. Meadows, 2008, chapter 6. ↩︎
- Three Years In, 10 July 2026 : building Theory as a technical organization, agents to map markets & research infrastructure to cover the ground of a firm several times our size. ↩︎
- Rails World 2026 Opening Keynote, David Heinemeier Hansson, Austin, 23 September 2026. He is co-owner & chief technology officer of 37signals. ↩︎
- Artemis Security：八个月内完成30,000个拉取请求，2026年8月平均每位工程师每天合并16个，高于5月份的6个和1月份的2个。↩︎
- Dan Shiebler，联合创始人兼首席技术官，Artemis Security：“我们平台上的每一行代码都是由AI智能体编写的。我们的工程师设计系统、设定约束，并审查输出。”↩︎
- pstack 完全指南，第1部分，Lauren Tan，2026年9月1日：“它让我能够以极高的信心每月向生产环境提交 2,000 个 PR。”她的验证技能示例采用 MIT 许可证。The New Stack 在《一位工程师每月向生产环境提交 2,000 个 PR》一文中对该工作流程进行了专题报道，2026年9月19日。↩︎
- 系统思维导论，Donella H. Meadows，2008年，第6章。↩︎
- 三年之际，2026年7月10日：将 Theory 构建为一家技术组织，利用智能体映射市场与研究基础设施，以覆盖相当于我们规模数倍的公司业务领域。↩︎
- Rails World 2026 开幕主题演讲，David Heinemeier Hansson，奥斯汀，2026年9月23日。他是 37signals 的联合创始人兼首席技术官。↩︎
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