OpenAI Codex Generated 90% of Its Own Code

Gergely Orosz got rare access to the engineering behind OpenAI’s Codex, and the details are worth sitting with. The team chose Rust over TypeScript and Go for the core agent, prioritizing correctness and minimal dependencies over faster iteration speed. Engineers on the Codex team don’t write most of the code anymore – they run four to eight parallel agents and review what comes back. The pull request volume from agent-generated code is already straining traditional review pipelines, a bottleneck that other teams running heavy agentic workflows at Anthropic and elsewhere have quietly started hitting too. OpenAI open-sourced the CLI and core agent, which makes Codex one of the few production-grade agent architectures you can actually inspect.