OpenAI says its Codex AI agent wrote every line of code in new product repository

OpenAI, the artificial intelligence company behind ChatGPT, says its Codex coding agent produced every line of code in a new product repository that grew to approximately one million lines over five months.

The company said the first commit landed in late August 2025 and the repository expanded to roughly 1,500 pull requests, covering application logic, tests, continuous integration (CI) pipelines, documentation, observability tooling and internal developer tools.

A small human team steered the agents rather than writing code directly, the company said, with three engineers initially guiding the work before the group grew to seven.

OpenAI said Codex opened and merged roughly three and a half pull requests per engineer per day while hundreds of internal users tested the product, with the company reporting that single Codex sessions worked on tasks for upwards of six hours.

The company said it structured the repository so agents could operate without extensive external context, describing the approach as giving "Codex a map, not a 1,000-page instruction manual."

The repository uses a short AGENTS.md file as a table of contents and a structured documentation directory as the system of record, with the team exposing user interface elements, logs, metrics and browser developer tools directly to agents.

OpenAI said agents generate not only product code but also tests, CI and release tooling, documentation, evaluation harnesses, review comments and repository management scripts.

The company described mechanical enforcement of coding standards through custom linters, structural tests and a recurring documentation maintenance agent that scans for outdated material.

Humans interact primarily through prompts and increasingly delegate review work from one agent to another, the company said, with an end-to-end capability set that includes reproducing bugs, implementing fixes, validating changes, opening pull requests, responding to feedback and merging code.

Related reading

• OpenAI to deploy ChatGPT on Pentagon's secure AI platform
• OpenAI tests ads in ChatGPT
• Google Photos adds AI-powered Ask button for image search and editing

OpenAI cautioned that the approach has worked through internal launch but said it is still learning how architectural coherence will evolve and where human judgment adds the most value.

The company added that the results depend heavily on the repository's specific structure and tooling, and should not be assumed to generalise without similar investment.

The Recap

• Codex agents wrote the repository with no human-written code.
• Repository reached about one million lines and fifteen hundred pull requests.
• Agents can reproduce bugs, implement fixes, and merge changes autonomously.