Scaling long-running autonomous coding (cursor.com)

Researchers have successfully scaled autonomous coding agents to work collaboratively on complex projects, such as building a web browser from scratch, generating over one million lines of code in just under a week. By implementing a dynamic coordination structure, the team transitioned from a flat system to a role-segregated pipeline, allowing planners to create tasks while workers focus solely on completing them. This new approach significantly improved efficiency and reduced coordination bottlenecks, though challenges such as agent risk aversion and the need for occasional resets remain. The implications for the AI/ML community are profound. This work highlights the potential for scaling coding tasks traditionally tackled by human teams, with the system’s success attributed to carefully designed roles and iterative experimentation with agent prompts. Importantly, the research suggests that model selection matters; specific models like GPT-5.2 outperform others in long-duration projects. As the team continues to refine their methods, their findings promise to enhance AI-assisted software development, potentially shaping future capabilities in the field.