OpenAI’s Codex, launched in 2025, is increasingly used to build and refine its own development tools. OpenAI staff say the vast majority of Codex’s work is now handled by Codex itself, creating a feedback loop that accelerates its improvement.
Codex operates as a cloud-based software engineering agent capable of writing features, fixing bugs, and proposing pull requests. It runs in sandboxed environments linked to a user’s code repository and can run multiple tasks in parallel. Codex is accessible through ChatGPT’s web interface, a CLI, and IDE extensions for VS Code, Cursor, and Windsurf. The Codex name traces back to a 2021 GPT-3–based model that powered GitHub Copilot’s tab completion, and staff have joked it may stand for “code execution.”
While Claude Code from Anthropic is a known competitor, OpenAI says the market is full of good ideas, and Embiricos declined to confirm direct influences. He noted that OpenAI had begun shipping Codex features internally before releasing the CLI version, which appeared after Anthropic’s tool.
A standout internal success story is the Sora Android app, developed by four engineers in 18 days with Codex helping to plan architecture, generate sub-plans for components, and implement them. Internally, Codex also helps monitor its own training runs and process user feedback to decide what to build next; engineers can assign Codex tasks via Linear, effectively integrating the AI agent into the team workflow.
OpenAI frames Codex as a teammate rather than a substitute for human workers, discussing concepts like vibe coding and vibe engineering to describe how humans stay in the loop. The company’s longer-term aim is to make coding agents useful to non-programmers as well, potentially turning Codex into a more general-purpose AI teammate rather than a tool limited to software engineers.
