Wilson Lin at Cursor coordinated hundreds of GPT-5.2 agents to build a web browser from scratch, running uninterrupted for one week. Over a million lines of Rust.
这个案例展示了AI系统的惊人规模和产出能力，协调数百个AI agent，一周内生成超过一百万行代码。然而，'远未达到生产质量'的评估也揭示了当前AI系统在复杂项目中的局限性，特别是在代码质量和系统架构方面。
Your AI agent writes every change into source code.
这一功能暗示了一种全新的开发范式，设计师的视觉编辑可以直接转化为生产级代码。这可能会显著减少前端开发中的手动编码工作，但也引发了关于AI生成代码质量和可维护性的重要问题。
Your AI agent writes every change into source code.
令人惊讶的是：这个工具将设计工作与代码生成完全分离，设计师可以直接在浏览器中进行视觉编辑，而AI代理会自动生成对应的CSS代码，这大大降低了前端开发的门槛，模糊了设计与开发之间的界限。
Without experience with compiler behavior, the agent couldn't have predicted which 'optimizations' the compiler would already handle.
这一观察揭示了AI代理在编译优化方面的局限性：代理无法准确预测编译器已经自动处理的优化。这表明AI代理需要更深入理解编译器行为和现代编译技术，以避免徒劳的优化尝试。这一发现对AI辅助编程系统的发展具有重要启示，强调了领域知识整合的重要性。
Modern clients are produced by a modern code generator, combined with hand-crafted functionality for some services.
The libraries in this repo are simple REST clients. These clients connect to HTTP/JSON REST endpoints and are automatically generated from service discovery documents. They support most API functionality, but their class interfaces are sometimes awkward.
Besides that ffscreencast can act as an ffmpeg command generator. Every available option can also just show the corresponding ffmpeg command instead of executing it. Non-ffmpeg commands, such as how the camera resolution is pulled and others can also be shown instead of being executed.
So transcriptor aims to do less, and impose the bare minimum of cognitive load needed to convert a REPL interaction into a test. The entire API is four functions:
You can find code to generate test fixtures