Cursor reveals its "bootstrapping" training method: using the old Composer to build the environment for the new model, resulting in a 14-point increase in Terminal-Bench.

According to Beating's monitoring, Cursor has revealed a training technique for its Composer series models: automatically building a runnable environment for the next generation of reinforcement learning (RL) using the previous generation model. When training Composer 2, Cursor uses Composer 1.5 to accomplish this task, calling it autoinstall. RL training requires a runnable code environment. If the environment is not set up properly, the model wastes tokens debugging and learns nothing; in extreme cases, the environment completely fails, wasting the entire training round's computational power. Autoinstall solves this problem in two steps: First, an agent reads the codebase documentation and configuration, generating 10 verification commands and expected outputs; second, another agent takes 3 of these commands and sets up the environment from scratch until the commands run successfully. The second step can be retried a maximum of 5 times; if all attempts fail, the environment is discarded. During the environment setup process, the agent proactively fills in missing dependencies: forging database tables, creating MinIO configurations to replace S3, starting Docker containers as sidecar services, and even generating placeholder images. The blog post uses the blockchain project celo-org/celo-monorepo as an example to demonstrate the entire process. After the agent failed to configure the environment in the first round, it created a mock user in the second round to bypass authentication and finally succeeded in the test. Composer 2 scored 61.7% on Terminal-Bench (a benchmark for testing the ability to set up a development environment for a model), nearly 14 percentage points higher than Composer 1.5's 47.9%. Cursor stated that it plans to involve the older version of Composer in more training stages in the future, including data preprocessing, runtime management, and architecture tuning.