Cursor has unveiled a novel 'autoinstall' training method for its Composer series models, leveraging previous-generation models to set up environments for reinforcement learning (RL). During the training of Composer 2, Composer 1.5 was used to automatically configure runnable environments, addressing issues of poorly configured setups that can waste computational resources. The process involves reading codebase documentation to propose verification commands and building environments until successful execution. This method improved Composer 2's performance on the Terminal-Bench benchmark to 61.7%, a significant increase from Composer 1.5's 47.9%. Cursor aims to further integrate older models into additional training stages.