A notable aspect of how we develop Composer is the way we use past versions of the model to improve the training process for future ones.
One of the clearest opportunities for this kind of bootstrapping is environment setup. RL training requires runnable environments, and if the environment is broken at the start, the model wastes tokens debugging setup instead of learning to solve problems. In the worst cases, a bad environment can make a problem unsolvable entirely, which ends up burning compute for no reward signal.
To address this, we built Composer autoinstall, a system that uses earlier Composer models to automatically create working RL environments from unconfigured repository checkouts. During training of the most recent version of the model, Composer 2, we used its predecessor, Composer 1.5, to manage this process. Beyond simply following step-by-step instructions, we found that modern coding models will go to great lengths to successfully configure, mock project dependencies, and test that setup is successful.
Better environments mean better training signal
Like many aspects of our model development, autoinstall is inspired by production Cursor systems. In Cursor cloud agents, we have a feature that automates the setup of cloud environments for users, to allow their agents to work on projects in a mock environment. Starting from a git checkout, the agent works to install packages, configure settings, and run basic checks to ensure that the code is running and stable. This allows future requests to start from the correct setup.
