OpenAI builds product with zero human-written code using Codex
Original: Harness engineering: Leveraging Codex in an agent-first world
Why This Matters
Demonstrates practical viability of AI-generated code at enterprise scale
OpenAI team built internal software product over 5 months using only Codex-generated code. Project reached 1 million lines across 1,500 pull requests with 3-7 engineers directing agents, achieving 10x faster development than manual coding.
OpenAI's engineering team conducted an experiment building a software product with zero manually-written code lines. Over five months starting August 2025, they used Codex agents to generate all code including application logic, tests, CI configuration, and documentation. The project reached approximately 1 million lines of code across 1,500 pull requests, with initial team of 3 engineers growing to 7, averaging 3.5 PRs per engineer daily. The product has hundreds of internal users including daily power users. Engineers' role shifted from coding to designing environments, specifying intent, and building feedback loops for reliable agent work. OpenAI estimates development was 10x faster than manual coding. The repository started empty with initial scaffold generated by Codex CLI using GPT-5, including the AGENTS.md file directing agent workflows.