Cursor CEO Michael Truell led an experiment where hundreds of AI agents, specifically GPT-5.2, collaborated to create a web browser from scratch. Over the course of a week, these agents produced more than 3 million lines of code. The project showcased the capabilities of GPT-5.2, which is designed for long-term autonomous tasks. However, initial attempts at self-coordination among the agents failed, leading to indecision. The breakthrough came when they assigned distinct roles: planners set tasks, workers completed them, and judges evaluated progress. This hierarchical structure resolved the coordination challenges. The newly created browser, while not on par with established ones like Chromium or WebKit, can render simple websites effectively. Truell noted that it handled issues autonomously, debugging itself as it progressed. Despite this achievement, there are significant concerns about the practicality of the code generated. Traditional browsers are built over years with tens of millions of lines of code, encompassing complex features like security protocols and extensions. The AI-generated browser lacks many of these essential components, making it far from production-ready. Developers pointed out difficulties in navigating the code and emphasized that human input was critical in defining the project's scope and goals. Sustainability poses another challenge. Writing extensive code is one thing, but maintaining it is another. Fixing bugs in the generated code could take longer than the initial coding phase. While this experiment demonstrates exciting potential in AI’s ability to assist in software development, it also highlights the limitations and necessary human oversight in creating complex software systems. Cursor plans to integrate the multi-agent coordination methods into its main product, indicating ongoing interest in refining this approach.