OpenAI’s coding agent Codex now writes the majority of its own code under developer supervision

OpenAI is using an AI coding agent and they also uses it to improve the agent itself. It is being revealed that its artificial intelligence coding agent, Codex, is now responsible for building most of its own software. The disclosure marks an important moment in how AI tools are being used to improve themselves.

According to Alexander Embiricos, who is product lead for Codex at OpenAI, the system is now “almost entirely being used to improve itself”. He shared this information in an interview with Ars Technica this week.

What is Codex and what does it do?

Codex is a cloud-based AI software engineering agent launched by OpenAI in May 2025. It is designed to help developers with tasks such as:

Writing new features

Fixing bugs

Reviewing code

Suggesting pull requests

The tool works in secure sandboxed environments connected to a user’s code repository and can run multiple tasks at the same time.

OpenAI offers Codex through:

ChatGPT’s web interface

A command-line tool (CLI)

IDE extensions for VS Code, Cursor and Windsurf

OpenAI says Codex is now deeply integrated into its own development process.

According to the company:

More than 92 per cent of OpenAI’s technical staff use Codex

Codex reviews every internal pull request

Engineers using Codex ship around 70 per cent more merged pull requests

Codex catches bugs before they reach production

At OpenAI DevDay 2025, the company revealed that Codex wrote 80 per cent of the Agent Builder tool in less than six weeks.

This information was shared by OpenAI during public events and interviews.

How Codex helps improve itself

Codex is powered by codex-1, a specialised version of OpenAI’s o3 reasoning model, trained using reinforcement learning on coding tasks.

OpenAI later released:

GPT-5-Codex in September 2025, designed for long, independent coding sessions

GPT-5.1-Codex-Max, released in early December via GitHub Copilot

According to OpenAI, Codex can now:

Monitor its own training runs

Review user feedback

Help decide what features to build next

OpenAI employees can assign tasks to Codex using the same tools they use for human colleagues, such as Linear and Slack.

Faster development with AI “teammates”

One example shared by OpenAI was the Sora Android app.

According to Embiricos: The app was built by four engineers It took 18 days to build and 28 days to ship. Codex helped plan the architecture and write code

OpenAI says Codex acts like a junior developer, working under human supervision. Engineers review all code before it is merged.

Competition in AI coding tools

OpenAI faces strong competition in the AI coding space from:

Anthropic’s Claude Code

GitHub Copilot

Google’s Gemini CLI

Startups like Cursor

On benchmarks, OpenAI says its latest models show improvement.

According to company data, GPT-5.2 scored 55.6 per cent on SWE-Bench Pro, higher than earlier versions and some competitors.

Will AI coding tools replace developers?

OpenAI says Codex is designed to support developers, not replace them.

According to Embiricos, humans remain in control:

Developers review all code

AI is used for speed and assistance

Final decisions stay with people

OpenAI believes coding agents could eventually help even non-programmers, but says human oversight will remain important.