Composer is a new agent model developed for software engineering, designed to significantly enhance coding speed and intelligence. It operates four times faster than similar models, achieving high-quality results in real-time coding scenarios. The model is trained on complex software engineering tasks within large codebases, using real-world production tools to tackle various challenges. This training approach aims to keep developers engaged and in the flow of coding, which was a key insight from the earlier Cursor Tab project. Built as a mixture-of-experts (MoE) model, Composer uses reinforcement learning (RL) to hone its capabilities in diverse development environments. Each training iteration involves the model addressing specific problems with the option to perform actions like code editing or executing terminal commands. The effectiveness of Composer is evaluated using Cursor Bench, an internal benchmark that assesses how well the model meets the needs of software developers, focusing on correctness and adherence to established coding practices. Significant infrastructure investments were required to train these large models effectively. Composer employs a custom training setup that utilizes PyTorch and Ray for asynchronous RL, allowing it to scale across thousands of NVIDIA GPUs while maintaining low communication costs. The model can access various tools for editing code and searching semantically, necessitating the management of hundreds of thousands of coding environments simultaneously in the cloud. The development team actively uses Composer for their own projects, indicating its practical utility in daily software development tasks.