OpenAI launches Codex: A cloud agent for software engineers that writes code, fixes bugs, and handles tasks in parallel

OpenAI just unveiled its latest engineering marvel: Codex, a cloud-based software engineering agent designed to write code, fix bugs, and assist developers by running multiple tasks at once—all in isolated environments tailored to your repo.

“Today, we are introducing Codex. It is a software engineering agent that runs in the cloud and does tasks for you, like writing a new feature or fixing a bug. You can run many tasks in parallel,” OpenAI CEO Sam Altman said on X.

What is Codex?

Codex is a research preview that gives developers an AI-powered teammate capable of autonomously handling engineering work, from writing new features and answering codebase questions to running tests and proposing pull requests. Each task Codex takes on runs in its own cloud sandbox, preloaded with your codebase.

“Codex is a cloud-based software engineering agent that can work on many tasks in parallel, powered by codex-1. Available to ChatGPT Pro, Team and Enterprise users today, Plus soon,” OpenAI states in a blog post.

It’s powered by Codex-1, a specialized version of OpenAI’s o3 model, trained through reinforcement learning on real-world programming tasks. The goal? Make Codex behave more like a disciplined, high-performing software engineer—clean code, precise task execution, and thorough testing.

How Codex Works

Codex is now available to ChatGPT Pro, Enterprise, and Team users via the ChatGPT sidebar. Just type a prompt and hit “Code” or ask a question about your codebase via “Ask.” Each request is handled in a separate, secure environment. Codex can read and edit files, run tests, check types, and even lint your code. Task completion time ranges from 1 to 30 minutes, depending on complexity.

Once done, Codex commits its changes and shares verifiable outputs—logs, test results, and more—so developers can trace every step. You can then request follow-ups, push a GitHub PR, or sync changes to your local dev setup.

To fine-tune Codex’s behavior, developers can add AGENTS.md files to their repo, like README.md, but with instructions specific to how Codex should navigate, test, and contribute to your codebase. This allows teams to enforce project-specific standards and make the most of Codex’s capabilities.

Performance and Benchmarks

Codex shows impressive performance in both internal and external benchmarks. On OpenAI’s curated set of real-world SWE tasks, codex-1 achieves pass@1 accuracy as high as 80%—significantly outperforming previous models like o1 and o4-mini. Even without AGENTS.MD configuration, Codex still performs reliably.

A Step Toward Trustworthy AI Coding Agents

Security and transparency were core to Codex’s design. Every action is traceable. If something goes wrong, such as a failed test, Codex lets you know explicitly. This traceability helps users catch issues early and ensures a higher bar for safety, especially as autonomous agents become more capable.

That said, OpenAI stresses the importance of human oversight: users should always manually review agent-generated code before merging or deploying it.

Codex Is Rolling Out Now

Codex is rolling out starting today for ChatGPT Pro, Enterprise, and Team plans, with support for Plus and Edu users coming soon.

As Altman put it, Codex isn’t just another coding assistant—it’s a parallelizable engineering agent. The dream of offloading tedious or complex software tasks to AI is edging closer to reality.