Cursor has launched its long-running agents research preview for Ultra, Teams, and Enterprise users. This development builds on previous research, specifically the creation of a web browser, revealing how frontier models struggled with complex, long-term tasks. In response, Cursor developed a custom harness that allows these agents to tackle larger projects more effectively. Early results show that long-running agents generated significantly larger pull requests (PRs) with merge rates similar to traditional agents. Participants in the preview used long-running agents to accomplish tasks that typically required more human oversight. Examples include building a new chat platform in 36 hours, creating a mobile app from a web app in 30 hours, and refactoring an authentication system in 25 hours. The agents excelled by proposing plans before execution and collaborating to maintain focus on complex tasks. Initial feedback indicates that these agents can work for extended periods, allowing users to engage in other work while projects progress. Internally, Cursor has tested these agents on various production tasks, such as optimizing a video renderer by migrating to Rust and implementing custom kernels. They also tackled network policy controls for sandboxed processes and secured sudo password prompting for the CLI. These efforts highlight the potential for long-running agents to automate more complex workflows with minimal human intervention, paving the way for future developments in self-driving codebases. Cursor is focused on enhancing collaboration between agents and improving the management of the increasing volume of generated code.