Developers at Cursor have introduced a novel system, dubbed Composer autoinstall, designed to bootstrap the AI model training process. This system leverages earlier versions of the Composer AI to automatically configure runnable environments from un-configured code repositories. The goal is to eliminate wasted computational resources spent on debugging setup issues, allowing the AI to focus on learning.
This bootstrapping approach is particularly crucial for Reinforcement Learning (RL) training, which demands stable, functional environments. A faulty setup can render problems unsolvable, leading to significant compute waste without any learning signal. Composer 2, the latest iteration, was trained using its predecessor, Composer 1.5, to manage this complex environment setup.
From Codebase to Capable Environment
The autoinstall system draws inspiration from Cursor's existing cloud agents, which automate development environment setup for users. Starting with a Git checkout, these agents install packages, configure settings, and perform basic checks to ensure code stability. For RL training, the stakes are higher, aiming to create a runnable mock environment from a codebase, enabling the AI to tackle future, unseen coding challenges.
Composer is trained with a comprehensive suite of tools, including linters, search functionalities, and sandboxed shell access. Effective environment setup is paramount; failure here directly impacts training efficiency and resource utilization.
A Two-Stage Process: Goal Setting and Execution
Autoinstall operates in two distinct phases. First, a "goal setting" agent receives the codebase and proposes a set of commands and expected outputs for a correctly configured environment. This agent analyzes documentation and existing files, attempting common installation and testing commands.
In the second stage, a separate Composer agent receives the initial environment state and a selection of target commands. This agent then explores the codebase, using tool calls to establish the necessary environment for the commands to execute successfully. Success is verified by running all target commands and comparing their output to the initial target descriptions.
The system is robust, allowing for multiple retries within the second stage. If an environment fails to meet satisfactory standards after five attempts, it is discarded. Composer autoinstall goes beyond basic setup, mocking missing files, creating placeholder assets, and even simulating database tables to achieve a complete environment.
Real-World Application: The Celo Project
To demonstrate its capabilities, autoinstall was applied to the complex Celo blockchain project. The process involved the agent navigating sparse documentation and sparse code to identify key installation commands, supplemented by web searches for crucial setup information.
The second stage proved challenging, requiring the agent to install additional dependencies like Foundry and mock an authentication flow. After an initial failure, the agent successfully created a mock user to run a minimal application, satisfying the testing requirements.
Bootstrapping the Next Generation of AI
Composer autoinstall represents a significant step in bootstrapping AI development processes. Composer 2 now demonstrates a marked improvement on Terminal-Bench, a benchmark for environment setup capabilities, scoring 61.7% compared to Composer 1.5's 47.9%. This advancement suggests future Composer versions will benefit from enhanced autoinstall capabilities, with prior instances likely playing roles in run management, data preprocessing, and architecture tuning.
