The article highlights experiments with autonomous coding agents designed to tackle complex projects that typically require significant human effort. The focus was on using hundreds of agents to complete a project that could take human teams months. During the tests, the agents collectively wrote over a million lines of code and handled trillions of tokens. Initial attempts at dynamic coordination faced challenges, such as agents getting stuck due to locking mechanisms and a lack of hierarchy, which led to risk-averse behavior and inefficiency. Switching to a more structured approach helped. They established distinct roles: planners create tasks and can spawn sub-planners, while workers focus exclusively on completing assigned tasks without worrying about coordination. This streamlined the process, allowing the agents to work on ambitious projects like building a web browser from scratch, which generated over a million lines of code in less than a week. Other experiments showed agents successfully migrating a codebase from Solid to React and significantly improving video rendering speed. The findings indicate that model choice is critical for long-running tasks. The GPT-5.2 model outperformed others in maintaining focus and implementing tasks accurately. The team learned that simplifying systems often leads to better outcomes, as adding complexity can create bottlenecks. Although the multi-agent coordination challenge remains, the results suggest that deploying many agents can lead to substantial progress on complex coding tasks.