OpenAI brings Codex into ChatGPT as a cloud coding agent
OpenAI's Codex research preview marked an important shift in how the company sees software engineering. In the official demo, Greg Brockman, Jerry Tworek, Joshua Ma, Hanson Wang, Thibault Sottiaux, Katy Shi and Andrey Mishchenko introduced Codex in ChatGPT as a remote software engineering agent that can run many tasks in parallel. The video was published by OpenAI on May 16, 2025, alongside the broader product announcement.
The central idea is simple but powerful: instead of asking an AI model for code in a chat window, a developer can delegate work to a cloud agent that has access to a repository, a configured environment and its own sandbox. Codex can answer questions about a codebase, fix bugs, write tests, propose changes and prepare pull requests. Each task runs separately, so users can start several jobs, continue with other work and return later to review the results.
According to the demo, the system initially rolled out to ChatGPT Pro, Enterprise and Team users, with Plus and Edu planned later. It was powered by Codex-1, a coding model optimized not only for benchmark performance but also for code people would actually want to merge: fewer unnecessary changes, better style and clearer summaries of what the agent did.
The demo also showed why this matters. Codex ran in isolated microVM sandboxes with filesystem, CPU, memory and network policies. It could read repository instructions, reproduce bugs, write regression tests, run linters and show proof of its testing work. That makes the interface closer to reviewing a junior engineer's pull request than copying code from a chatbot.
For developers and companies, the bigger message was about workflow. AI coding is moving from autocomplete and pair programming toward asynchronous delegation. That creates leverage, but it also changes the job: humans still need to define goals, configure environments, review diffs, check tests and decide what should be merged. Codex in ChatGPT was an early public signal that software teams would soon manage fleets of coding agents, not just prompts.