OpenAI Codex agent writes code

Code-writing AI has been inching toward this for a while. First it completed lines. Then it suggested functions. Now OpenAI is pushing Codex as something closer to a software engineering agent — a system that can take a task, work inside a sandboxed environment, run tests, and hand back a pull request instead of just a snippet. That is the real shift here. The product is no longer “help me type faster.” It is “go do the work, then show me the diff.” (developers.openai.com) ### What is Codex now? Codex is OpenAI’s coding agent inside ChatGPT and related developer tools. In the cloud version, you connect a GitHub repository, give it a task, and it spins up its own environment to read, edit, and run code. OpenAI’s current framing is broad: build features, fix bugs, refactor code, answer repo questions, and propose pull requests for review. That mak(developers.openai.com)e formatting. (developers.openai.com) ### Why does the cloud sandbox matter? Because the hard part was never just generating code. The hard part was doing software work safely and repeatably. Codex cloud runs each task in an isolated environment rather than on your laptop, and OpenAI says the agent phase is offline by default unless internet access is explicitly enabled. Setup can fetch dependencies, but the actua(developers.openai.com)g to make “agent that can run commands” feel less like giving root access to a stochastic intern. (developers.openai.com) ### What does it actually do in GitHub? Two things. First, it can create pull requests from work done in the cloud after you connect GitHub. Second, OpenAI now offers direct GitHub integrations for review workflows. You can tag `@codex review` on a pull request and get a standard GitHub code review back, and teams can also enable automatic reviews. So Codex (developers.openai.com)e human engineers already coordinate changes. (developers.openai.com) ### How does this reach CI/CD? OpenAI also ships a Codex GitHub Action. That lets teams trigger Codex from GitHub events, run review or migration tasks in workflows, and post the results back to pull requests. The docs are pretty explicit about the use cases — automate feedback on PRs, gate changes with Codex-driven checks, and run repeatable tasks in CI pipelines. That is a b(developers.openai.com)uild-system plumbing. (developers.openai.com) ### Is this one agent or a bunch of tools? OpenAI is pitching it as one agent across surfaces — terminal, IDE, cloud, GitHub, even phone — tied together by the same account and state. There is also support for parallel work and explicit subagents, which means one coding task can branch into multiple smaller ones when you ask for it. The important idea is continuity: st(developers.openai.com) back. (5codex.com) ### What is the model underneath? The current official model line is GPT‑5.2‑Codex, not GPT‑5.5. OpenAI describes it as a GPT‑5.2 variant optimized for agentic coding, with better long-horizon work, stronger refactors and migrations, improved Windows performance, and stronger cybersecurity capability. So if you saw chatter about a “GPT‑5.5 Codex agent,” the safer reading is that the product story is real, but(5codex.com)enAI’s published release pages. (openai.com) ### What is the catch? The catch is trust. A coding agent that can edit files, run package managers, and touch production-adjacent workflows needs guardrails. OpenAI’s answer is sandbox modes, approval policies, network controls, and secrets that are removed before the cloud agent phase starts. That does not make mistakes impossible. But it does show where the market is going: (openai.com) cabinet.” (developers.openai.com) ### Bottom line Codex matters because it turns code generation into delegated execution. The interesting part is not that it writes code. Lots of systems do that. The interesting part is that OpenAI is wiring it into repositories, tests, reviews, and CI so the output arrives as work product — a patch, a review, a PR — instead of just text. (developers.openai.com)