Composer autoinstall: how older model versions help train newer ones

Cursor has introduced Composer autoinstall, a system that uses earlier versions of the Composer model to automatically prepare environments for reinforcement learning. During the development of Composer 2, the team used version 1.5 to manage this process. The idea builds on Cursor's experience with its cloud agents, but applies it to RL training of the models themselves.

Why broken environments kill RL training

RL training requires working environments. If a project won't compile, dependencies won't install or a configuration refuses to start, the model spends tokens on debugging instead of learning to solve real programming problems. In the worst cases, a broken environment makes the task unsolvable altogether: the model receives no reward signal and simply burns compute. That is expensive and inefficient.

The two-stage bootstrap process

Autoinstall works on a simple but clever scheme.

Stage 1: a scout agent defines the goal. The first model version (Composer 1.5) is given a repository in a fixed state. It must propose 10 commands, along with a high-level description of their output if the environment is set up correctly. The model reads the README and Makefile, tries language-specific commands (`uv`, `npm install`, `clippy`, `pytest`) and explores the project structure. The result is a list of setup commands, tests and launch scripts.

Stage 2: a second agent implements it. The second version (Composer 2) receives the project's initial state plus three target commands chosen from the proposed ten. Its task is to call tools (search, compilation, linter), explore the code and configure the environment so that all three commands run and their output matches the description from stage 1. If it doesn't match, the process repeats. After five failed attempts, the environment is discarded.

• The model explores the code and runs search tools
• Installs dependencies through a package manager
• Performs configuration (settings, environment variables)
• Checks the output against the target description
• Repeats until it succeeds or hits the attempt limit

How the model works around missing components

Composer will go to great lengths to get an environment working. The model mocks missing files, creates placeholder images and even fake database tables. If a project needs cloud services such as S3 or sidecar containers, Composer builds equivalents: MinIO configs for S3, Docker containers for services. For long-running processes, the system generates a startup script that launches these components at the start of the RL session.

"Modern language models will go to great lengths to successfully set up an environment, mock dependencies and test that the setup works," the

Cursor team says.

What this means for the future

The idea is simple but significant. Composer uses its own older version as a helper to prepare the working foundation for the new one. This not only saves compute but also improves the signal for reinforcement learning. Each new model version now stands on the shoulders of its predecessors. It is reasonable to expect that this kind of bootstrap will become standard in training large language models.