Cursor discloses the "bootstrap" training method: using the old Composer to set up the environment for the new model, Terminal-Bench increases by 14 points

According to Beating Monitoring, Cursor revealed a training trick for the Composer series models: using the previous generation model to automatically build a runnable environment for reinforcement learning (RL) of the next generation. When training Composer 2, Cursor used Composer 1.5 to accomplish this task, calling it autoinstall.

RL training requires a runnable code environment. If the environment is poorly set up, the model wastes tokens on bug fixing and learns nothing; in extreme cases, if the environment cannot run at all, the entire training compute is wasted. autoinstall solves this problem in two steps: first, an agent reads the codebase documentation and configuration, proposing 10 validation 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 retries up to 5 times; if all fail, the environment is discarded.

During environment setup, the agent actively fills in missing dependencies: fabricating database tables, creating MinIO configurations as a substitute for S3, launching 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 first environment setup failure, the agent in the second round creates mock users to bypass authentication, ultimately passing the test.

Composer 2 scored 61.7% on Terminal-Bench (a benchmark for testing the ability to build development environments for models), nearly 14 percentage points higher than Composer 1.5’s 47.9%. Cursor states that future plans include involving the older Composer in more training stages, including data preprocessing, runtime management, and architecture tuning.