Cursor has revealed a clever training trick for its Composer series of coding models: using an older version of the model to automatically set up working reinforcement learning environments for the next generation. When training Composer 2, the company used Composer 1.5 to do this task, calling the technique "autoinstall."
RL training requires a working code environment. If the environment isn't set up correctly, the model wastes tokens debugging without learning anything. In the worst case, the environment fails entirely, burning the entire training compute. Autoinstall solves this in two steps: First, an agent reads the codebase's documentation and configuration, then generates 10 verification commands along with their expected outputs. Second, another agent takes three of those commands and sets up the environment from scratch until the commands run successfully. This second step retries up to five times; if all fail, that environment is discarded.
When setting up the environment, the agent proactively fills in missing dependencies: faking database tables, creating MinIO configurations to replace S3, starting Docker containers as sidecar services, and even generating placeholder images. For example, with the blockchain project celo-org/celo-monorepo, after failing in the first round, the agent created mock users to bypass authentication in the second round, finally getting the tests to pass.
After training, Composer 2 scored 61.7% on Terminal-Bench (a benchmark for testing a model's ability to set up development environments), compared to Composer 1.5's 47.9%. Cursor believes this means using Composer 2 for autoinstall would produce even better results, creating a positive feedback loop. The company also plans to have older Composer versions participate in more training stages, including data preprocessing, runtime management, and architecture tuning.