835页文档。没有源码。没有测试用例。不能上网查资料。
给你这样一份SQLite的官方手册，让你从零开始重写一个功能完整的SQLite——你觉得你需要多久？一个团队要多久？
Cursor的答案是：一群Agent，几天时间。
InfoQ等多家媒体报道，Cursor团队最近完成了一项惊人的实验：用多Agent系统，仅凭一份835页的SQLite参考手册，从零开始复刻了一个功能完整的SQLite实现。没有给Agent看源码，没有给测试用例，没有让它们联网搜索——就靠读文档、写代码、调试、修复，Agent们自己把整个数据库系统搭了起来。
从补全到创造：AI编程的能力跃迁
AI编程工具的能力，在过去三年里经历了三次跃迁。
第一代是代码补全。GitHub Copilot刚出来的时候，大家最惊艳的是“写个注释它就能补全后面几行代码”。那时候AI编程的定位是“高级自动补全”——你告诉它要干什么，它帮你填细节。人还是主导者，AI只是助手。
第二代是功能实现。Cursor、Claude Code这些工具出现后，你可以说“帮我写一个用户登录系统”，它能生成一整套代码——前端、后端、数据库、测试都有。这时候AI已经能独立完成一个完整的功能模块了，但前提是你得把需求说清楚，而且它写出来的代码你得能看懂、能改。
第三代就是现在——从零构建复杂系统。重造SQLite这个实验的意义在于：AI已经能从一份自然语言（或者说技术文档）描述出发，自主构建一个有几十万行代码的复杂系统了。它不只是“填空”，不只是“做模块”——它能从零开始，自己读文档、自己理架构、自己写代码、自己调试、自己修复bug。
这中间的差别是什么？是“自主性”。以前的AI编程，你是项目经理，AI是程序员——你下指令，它执行。现在的AI编程，你更像产品经理——你告诉它目标（“我要一个能跑的SQLite”），它自己拆任务、自己排优先级、自己解决遇到的问题。
开发者会被取代吗？别问错了问题
每次AI编程能力进步，都会有人问“程序员会不会失业”。这个问题问错了方向。
真正的问题不是“程序员会不会被取代”，而是“程序员的工作内容会发生什么变化”。
想想看：SQLite这个实验里，Agent们是“从零复刻”，不是“从零发明”。它们有835页的手册作为明确的目标——知道要做什么、做成什么样。但在真实的软件开发中，最难的部分从来不是“怎么写代码”，而是“写什么代码、为什么写、写给谁用、怎么赚钱”。需求定义、产品决策、架构权衡、用户体验、商业价值——这些才是真正需要人来拍板的东西。
AI编程越强大，“写代码”这件事的价值就越低，“定义问题”这件事的价值就越高。未来的开发者，可能不需要自己写每一行代码，但需要更深刻地理解业务、更精准地定义需求、更敏锐地判断技术选型的利弊。你的价值不在于你能写多少行代码，而在于你能提出什么样的问题、做出什么样的判断。
当然，这个实验也有它的局限性。复刻SQLite是一个“有标准答案”的问题——手册上写得明明白白，Agent只需要照着做。但真实世界的软件开发，大部分问题没有标准答案，需要在不确定中做决策、需要在约束中找平衡、需要和人打交道。这些能力，AI还差得远。
但方向是明确的：AI编程正在从“帮你写代码”走向“帮你建系统”。能适应这个变化的开发者，会拥有前所未有的生产力；不能适应的，确实会被淘汰。但被淘汰的原因不是AI太强，而是人没有进化。
明天见。
835 pages of documentation. No source code. No test cases. No internet access.
Given just the official SQLite manual, how long would it take you to rewrite a fully functional SQLite from scratch? How long would a team need?
Cursor's answer: a swarm of agents, a few days.
Multiple outlets including InfoQ report that the Cursor team recently completed a stunning experiment: using a multi-agent system, they rebuilt a fully functional SQLite implementation from scratch using only an 835-page SQLite reference manual. No source code shown to the agents, no test cases provided, no internet search allowed — just by reading documentation, writing code, debugging, and fixing, the agents built the entire database system themselves.
From Autocomplete to Creation: The Leap in AI Coding Capability
AI coding tool capabilities have gone through three leaps in the past three years.
Generation one was code autocomplete. When GitHub Copilot first came out, everyone was amazed by “write a comment and it autocompletes the next few lines.” Back then, AI coding was positioned as “advanced autocomplete” — you tell it what to do, it fills in the details. Humans were still in charge; AI was just an assistant.
Generation two was feature implementation. With tools like Cursor and Claude Code, you can say “build me a user login system” and it generates a whole stack — frontend, backend, database, tests. By this point, AI can independently complete an entire functional module. But the prerequisite is you have to explain the requirements clearly, and you have to be able to understand and modify the code it produces.
Generation three is now — building complex systems from zero. The significance of the SQLite rebuild experiment is this: AI can now start from a natural language (or rather, technical documentation) description and autonomously construct a complex system with hundreds of thousands of lines of code. It's not just “filling in blanks” anymore, not just “building modules” — it can start from zero, read the docs itself, figure out the architecture itself, write the code itself, debug itself, fix bugs itself.
What's the difference? Autonomy. In earlier AI coding, you were the project manager and AI was the programmer — you gave orders, it executed. In AI coding today, you're more like a product manager — you tell it the goal (“I want a working SQLite”) and it breaks down tasks itself, prioritizes itself, solves problems it encounters itself.
Will Developers Be Replaced? You're Asking the Wrong Question
Every time AI coding capability advances, people ask “will programmers lose their jobs?” It's the wrong question.
The real question isn't “will programmers be replaced” — it's “how will developers' work change?”
Think about it: in the SQLite experiment, the agents were “rebuilding from scratch,” not “inventing from scratch.” They had an 835-page manual as a clear target — they knew what to build and what it should look like. But in real software development, the hardest part is never “how to write code” — it's “what code to write, why write it, who is it for, how does it make money?” Requirements definition, product decisions, architectural tradeoffs, user experience, business value — these are the things that really need human judgment.
The more powerful AI coding becomes, the lower the value of “writing code” becomes, and the higher the value of “defining the problem” becomes. The developers of the future might not need to write every line of code themselves, but they'll need deeper business understanding, more precise requirements definition, and sharper judgment on the pros and cons of technology choices. Your value isn't in how many lines of code you can write — it's in what kinds of questions you can ask and what kinds of judgments you can make.
Of course, this experiment has its limitations. Rebuilding SQLite is a problem with a “standard answer” — the manual spells everything out, the agents just need to follow along. But most problems in real-world software development don't have standard answers. They require decision-making under uncertainty, finding balance within constraints, dealing with people. These are capabilities AI is still very far from.
But the direction is clear: AI coding is moving from “helping you write code” to “helping you build systems.” Developers who adapt to this change will have unprecedented productivity. Those who don't adapt will indeed be left behind — but not because AI is too strong, but because humans didn't evolve.
See you tomorrow.
AI编程越强大，写代码的价值越低，定义问题的价值越高。你的价值不在于能写多少行代码，而在于能提出什么样的问题。
—— Dawn Vision编辑部
The more powerful AI coding becomes, the lower the value of 'writing code' — and the higher the value of 'defining the problem.' Your value isn't in how many lines you can write, but in what questions you can ask.
—— The Dawn Vision Editorial Desk
Cursor · SQLite · AI coding · multi-agent · autonomous programming · 835-page manual · build from zero · developer role
