168小时AI狂写300万行代码造出浏览器！Cursor公开数百个智能体自主协作方案

Core Insights - The article discusses a groundbreaking experiment by Cursor, where hundreds of AI agents collaboratively developed a usable web browser from scratch, producing over 3 million lines of code [2][3]. Group 1: Experiment Overview - The project, codenamed FastRender, resulted in a browser with a rendering engine written in Rust and a custom JavaScript virtual machine [2]. - The browser is described as "barely usable," with performance significantly lagging behind established browsers like Chrome, but it can render Google's homepage correctly [3][4]. Group 2: AI Model Utilization - The success of the experiment relied on OpenAI's GPT-5.2-Codex, which is designed for complex software engineering tasks and can autonomously plan and execute coding tasks [5][6]. - GPT-5.2-Codex incorporates a technique called "Context Compaction," enhancing its ability to maintain logical consistency while handling large codebases [8]. Group 3: Multi-Agent Collaboration - Cursor developed a multi-agent collaboration architecture to enable hundreds of AI agents to work simultaneously without conflicts [12][18]. - Initial attempts at a flat collaboration model led to significant inefficiencies, prompting a shift to a hierarchical structure with planners, workers, and judges to streamline the process [15][18]. Group 4: Insights and Challenges - The experiment revealed that the general GPT-5.2 model outperformed the specialized GPT-5.1-Codex in long-term autonomous tasks, while other models like Claude Opus 4.5 were better suited for interactive scenarios [21]. - The design of prompts was found to be more critical than the model itself, emphasizing the need for extensive trial and error to guide AI agents effectively [22]. Group 5: Future Implications - The experiment sparked significant industry discussion, with predictions that the marginal cost of software development could approach zero as token costs decline [25]. - Despite existing challenges, such as planning responsiveness and agent overactivity, the experiment demonstrated the feasibility of scaling autonomous coding capabilities through increased agent numbers [29].