OpenAI's Codex Learns Tasks by Watching You Once

OpenAI's Codex agent can now learn a task by watching you do it a single time. A new feature called Record & Replay lets you demonstrate a workflow on screen once; Codex captures the steps, converts them into a reusable "skill," and from then on can run the whole thing on its own without you walking it through again.

The launch example is mundane on purpose: uploading a video to YouTube, complete with its thumbnail, description and subtitles. Do it once with Codex watching, and the next upload becomes a one-click replay. The feature shipped in Codex version 26.616, alongside bulk actions for automation history and the ability to hand a thread off between your local machine and a remote one.

The catch: where and how it runs

There are real fences around it. Record & Replay lives in the Codex app on macOS, requires the agent's Computer Use mode to be switched on, and needs a paid ChatGPT account to do anything useful — the app itself is free to download. As of the June 20 announcement it is not available in the EU, the UK or Switzerland, though Computer Use itself opened up in the EU on June 16, so the regional gap may narrow.

If you've used agents before, you know the usual failure mode: you spend so long explaining a task in words that you might as well have done it yourself. Recording sidesteps that. You're not writing a spec, you're giving a demo — and a demo carries the small, fiddly details that are painful to describe but easy to show.

Why it matters for you

Most of the boring work on a computer is the same handful of steps repeated: rename and sort files, fill the same form, pull numbers into the same report, post content to the same three places. Those are exactly the tasks that are tedious to describe in a prompt but trivial to demonstrate. Record once, replay forever is a much lower bar than learning to write good instructions.

It also changes who can automate things. Turning a workflow into a script used to need someone who could code or at least wrangle a no-code tool. Showing the computer what to do, by just doing it, is a skill everyone already has. That's the quiet significance here — the interface for automation is shifting from language to demonstration.

The honest caveat: an agent that replays your clicks is only as reliable as the world staying still. The moment a button moves or a page redesigns, a recorded skill can break in ways a human would shrug off. Treat replays like a junior assistant, not a finished machine — useful, fast, and in need of a glance before you trust the output.