Introducing Codex

OpenAI has launched Codex, a cloud-based agent capable of handling software engineering tasks in parallel, as a research preview. Codex performs tasks such as implementing features, answering questions about codebases, fixing bugs, and proposing Pull Requests, with each task running in its own isolated cloud sandbox environment preloaded with the repository.

Codex uses codex-1, a model optimized for software engineering based on OpenAI o3. Through reinforcement learning on real coding tasks across diverse environments, it reflects human coding style and PR preferences, and has the ability to run iteratively until tests pass.

Key features and characteristics are as follows:

• Accessible via the ChatGPT sidebar, where users can assign tasks with 'Code' or ask questions about the codebase with 'Ask'.
• Can read and edit files, as well as run various commands including test harnesses, linters, and type checkers.
• An AGENTS.md file can be used to guide standard project practices, test commands, and how to navigate the codebase.

After a task is completed, the execution process can be transparently reviewed through terminal logs and test output. Users can review the results before integrating the changes or creating a GitHub Pull Request. It is currently being rolled out sequentially to ChatGPT Pro, Business, Enterprise, and Plus users.