How Codex Is Built
3 days ago
- #Software Engineering
- #OpenAI
- #AI Development
- OpenAI's Codex is a multi-agent coding assistant used by over a million developers weekly, with a 5x increase in usage since January.
- The Codex desktop app for macOS was launched in February, described by CEO Sam Altman as OpenAI's 'most loved internal product'.
- GPT-5.3-Codex is the first model that helped create itself, showcasing advancements in AI self-improvement.
- Codex was built with Rust for performance, correctness, and engineering quality, contrasting with Claude Code's TypeScript approach.
- The core agent and CLI of Codex are open source, emphasizing transparency and community involvement.
- Codex operates on a state machine model with an agent loop that includes prompt assembly, inference, response streaming, and tool calls.
- Safety measures include sandboxing to restrict network and filesystem access by default, ensuring secure usage.
- Over 90% of Codex's code is generated by Codex itself, highlighting its self-referential development capabilities.
- Engineers on the Codex team manage multiple parallel agents for tasks like feature implementation, code review, and bug fixes.
- Agent Skills extend Codex's capabilities, with over 100 internal skills like security best-practices and Datadog integration.
- Tiered code review involves AI reviews for non-critical code and mandatory human reviews for key components.
- Unique engineering practices include AGENTS.md for AI navigation, structured code for agent success, and Codex testing itself.
- New engineers onboard by pairing with experienced team members, emphasizing rapid adaptation to Codex's development style.
- OpenAI employees have unlimited access to Codex, fostering extensive use and innovation within the company.
- Research at OpenAI involves practical applications, like SQ Mah's transition from software engineering to research via the Vesuvius Challenge.