In a significant development for artificial intelligence, OpenAI has revealed the extent to which it uses its own AI coding agent, Codex, to develop and improve the tool itself. Company insiders report that the vast majority of the Codex software is now written by the AI, creating a powerful recursive loop where the technology accelerates its own evolution.
This self-improvement cycle is reshaping how OpenAI's engineers work, with the company treating the AI agent as an integrated member of its development team. The approach has led to dramatic efficiency gains, such as building the Sora Android app from scratch in just 18 days with a small team of four engineers.
Key Takeaways
- OpenAI's AI coding agent, Codex, is now predominantly used to write its own code, creating a self-improvement cycle.
- The company's engineers treat Codex as a "teammate," assigning it tasks through project management tools like Linear.
- Using Codex, a team of four OpenAI engineers built the Sora Android app in 18 days.
- Internal adoption of Codex has surged, with usage jumping 20-fold after the release of an interactive command-line interface.
- While OpenAI touts productivity gains, some independent research suggests AI coding tools can slow down experienced developers on complex projects.
The AI That Codes Itself
The concept of a tool building a better version of itself is not new in technology. Early computer chips, designed by hand, enabled the creation of software that could design far more complex circuits. OpenAI is now applying this principle to software engineering on an unprecedented scale.
"I think the vast majority of Codex is built by Codex, so it’s almost entirely just being used to improve itself," explained Alexander Embiricos, product lead for Codex at OpenAI. This feedback loop means each new version of the AI is more capable, partly because its predecessor helped build it.
Codex, first introduced as a research preview in May 2025, functions as a software engineering agent. It can write new features, fix bugs, and propose code changes within a developer's existing projects. The tool is available through multiple interfaces, including ChatGPT, a command-line interface (CLI), and extensions for popular code editors.
From Tab Completion to Teammate
The "Codex" name has a history at OpenAI, dating back to a 2021 model that powered the initial version of GitHub Copilot. Embiricos noted that the original tool was a "'wow' moment for AI" for many developers, showcasing the potential for AI to understand context and accelerate coding tasks. The new agentic version of Codex represents a significant evolution from simple code completion to a more autonomous partner in the development process.
The company's own engineers are among its most active users. Embiricos confirmed that the majority of OpenAI's engineering team uses Codex regularly. Crucially, they use the same open-source version available to the public. "The version of Codex that we use is literally the open source repo. We don’t have a different repo that features go in," he stated.
A New Member of the Team
OpenAI's internal workflow has been fundamentally altered by Codex's integration. The AI is no longer just a tool but is treated as a junior developer and a member of the team. This integration extends to the company's project management and communication platforms.
Ed Bayes, a designer on the Codex team, described the process. "You can add Codex, and you can basically assign issues to Codex now," he said. "Codex is literally a teammate in your workspace."
This means an engineer or designer can assign a bug fix or feature request to the AI agent directly within a tool like Linear, just as they would with a human colleague. The AI can then create a pull request with the necessary code, which the team can review and approve.
"It’s basically approximating this kind of coworker and showing up wherever you work," Bayes added, highlighting how the AI meets developers in their existing workflows, such as Slack channels.
This approach has empowered employees beyond traditional engineering roles. Bayes, a designer, now uses Codex to build functional prototypes himself, rather than handing off specifications to an engineering team. "It kind of gives you more leverage. It enables you to work across the stack and basically be able to do more things," he commented.
Accelerating Development and Raising Questions
The most striking example of Codex's impact is the development of the Sora Android application. According to Embiricos, the project's speed was a direct result of using the AI agent.
Sora Android App: By the Numbers
- Engineers: 4
- Time to Build: 18 days
- Time to App Store: 28 days total
The team leveraged existing iOS and server-side components, using Codex to plan the Android client's architecture and implement its various components.
Despite these internal successes, the broader impact of AI on developer productivity remains a subject of debate. An independent study by METR published in July found that experienced developers working on mature, complex codebases were actually 19 percent slower when using AI coding assistants. The researchers did note, however, that AI tools might perform better on new or simpler projects.
OpenAI's team distinguishes between different ways of using the tool. Embiricos contrasted "vibe coding," where a developer might accept AI-generated code without much review, with "vibe engineering," a more collaborative process. "You ask Codex to work on that, maybe you even ask for a plan first. Go back and forth, iterate on the plan, and then you’re in the loop with the model and carefully reviewing its code," he said.
The Future of Coding and Competition
OpenAI is not alone in the race to build powerful AI coding agents. The market is increasingly crowded, with major players and startups vying for dominance.
Key Competitors in AI Coding:
- Anthropic: Claude Code
- Google: Gemini CLI
- Mistral AI: Devstral 2 and Mistral Vibe
- Startups: Cursor, which has built a dedicated AI-native IDE.
The focus on coding as a primary application for large language models (LLMs) is intentional. "We have absolutely noticed that coding is both a place where agents are gonna get good really fast and there’s a lot of economic value," Embiricos confirmed.
The question of job displacement is never far from the conversation. Bayes and Embiricos maintain that Codex is an amplifier of human capability, not a replacement. They emphasize that a human remains in the loop to review code and provide direction. At OpenAI, the introduction of Codex has not led to a reduction in engineering headcount.
Looking forward, the vision extends beyond professional developers. Embiricos believes the technology being pioneered with Codex will eventually become a more general-purpose agent, accessible to everyone. "All humanity is not gonna open an IDE or even know what a terminal is," he said. "We’re building a coding agent right now that’s just for software engineers, but we think of the shape of what we’re building as really something that will be useful to be a more general agent."