OpenAI Codex: Cloud-Based Coding Agent Powered by codex-1
OpenAI has released Codex, a cloud-based coding agent designed for software engineering. Codex is powered by the codex-1 model, a version of OpenAI o3 optimized specifically for coding tasks, trained via reinforcement learning on real-world coding tasks to mirror human style, follow instructions precisely, and iteratively test code until it passes.
Technical Architecture and Execution Environment
Codex operates within isolated cloud containers to ensure security and controlled execution. Each agent runs in its own container with no internet access during the model trajectory. The environment is preloaded with the user's codebase and a development environment defined by the user, including all necessary dependencies, configuration, and tooling.
Within this isolated environment, Codex has the following capabilities:
- File System Access: The agent can read and edit files within the container.
- Command Execution: Codex can execute terminal commands, including running tests, linters, and type checkers to verify its own work.
Workflow and Verifiability
Codex is designed to provide verifiable evidence of its actions. The model provides citations of terminal logs and files to allow users to validate the work performed by the agent.
Once a task is complete, the user has three primary options for integrating the results:
- Inspection: Users can review the results of the agent's work.
- Refinement: Users can request further changes or refinements to the model's output.
- Export: Users can export the generated diff as a GitHub pull request or copy the code for local testing and development.
Relationship to o3 and o4-mini
This announcement serves as an addendum to the system cards for OpenAI o3 and o4-mini, providing specific details on the Codex agent's capabilities and its technical implementation using the codex-1 model.