Gergely Orosz · 12 min read · Gen AI Tools
How Codex is built
Read the original on newsletter.pragmaticengineer.com →
Codex, OpenAI's multi-agent coding assistant, has reached over a million weekly users and gained momentum with a macOS desktop app and GPT-5.3-Codex. A deep-dive with OpenAI engineers reveals how Codex is built—from Rust-based architecture and agent loops to its self-writing capabilities and the engineering practices that sustain rapid iteration.
Key Takeaways
- Codex reportedly writes more than 90% of the app's code, illustrating its capability to self-construct significant portions of the product.
- OpenAI favors Rust and open-source components, leveraging a multi-agent, agent-loop design to power Codex.
- Engineering practices include tiered code reviews, Codex self-testing, rapid onboarding via pairing, and data-driven evals with continuous dogfooding and A/B testing.
- Roadmap aspects highlight ongoing model evolution (e.g., GPT-5.x) and broad adoption across the developer ecosystem.
Excerpt
More than a million developers use OpenAI’s multi-agent coding assistant every week. Named Codex, usage has increased 5x since the start of January, the team tells me. In the first week of February, OpenAI launched the Codex desktop app, a macOS application that CEO Sam Altman calls “the most loved internal product we’ve ever had”. A few days later, OpenAI shipped GPT-5.3-Codex, which they describe as the first model that helped create itself. Personally, I’ve been warming up to Codex since doing an interview for The Pragmatic Engineer podcast with Peter Steinberger , the creator of OpenClaw, in which he…