OpenAI’s Codex Can Learn a Task After Just One Example

OpenAI’s Codex agent can now learn how to perform a task after seeing you do it only once. They’ve added a new Record & Replay feature which allows you to show Codex how to perform a workflow and then it learns it, creating a “reusable skill” that it can execute automatically in future, without further instruction.

They’ve kicked this off with a pretty banal use case — posting a video to YouTube, including a thumbnail image, a description and subtitles. Once you’ve shown Codex how to do this once, you can simply click the ‘replay’ button to automate the process again in future. This is available in Codex version 26.616, which also includes bulk operations for automation history, and moving threads between your local and remote systems.

1The catch: where and how it runs

This has some pretty significant limitations. Firstly, this only works via the Codex app on macOS and you need to enable Codex’s Computer Use mode to get it working. Secondly, you need to have a paid ChatGPT subscription to make any use of this (the app itself is free). At the time of writing (June 20) this feature isn’t available in the EU, UK or Switzerland, but Computer Use mode was launched in the EU on June 16th, so hopefully that will change soon.

For those of you familiar with agents, you’ll know the pitfalls of explaining a task in words — it’s often quicker to just do it yourself. By recording a workflow instead, you’re essentially demonstrating rather than documenting, which means that you’re capturing all of those little nuances that are hard to articulate but easy to demonstrate.

2Why it matters for you

A lot of the repetitive processes we do on our computers involve the same set of steps, whether that’s renaming and organising files, filling out forms, aggregating data into reports, or sharing content across multiple channels. These sorts of workflows are hard to explain in a prompt but easy to demonstrate, so recording once and replaying many times is going to be easier than learning how to document effectively.

This also opens up automation to more people. In the past, you needed a developer or someone with experience using a no-code platform to convert your workflow into a script, but now anyone can show the computer what to do, by doing it themselves. This is a subtle but important shift in the way we interact with automation tools — from language to demonstration.

That said, a replayable agent is only as robust as the real world. If a button position changes or a website gets redesigned, your recorded skills may stop working in unexpected ways. Think of replays as an intern, not a robot, they’re incredibly useful but they need supervision before you can rely on their results.