Codex Can Now Learn Your Workflow by Watching Once — What That Means for Your Work

TL;DR. On June 18, 2026, OpenAI released Record & Replay for the Codex macOS app: show it a workflow once and it saves a reusable Skill it can run automatically. It requires a paid ChatGPT account, Computer Use enabled, and a Mac. It is not yet available in the EU, UK, or Switzerland. Here's what that actually means for your work.

The core promise of AI workflow tools has always been "do the repetitive parts for you." Record & Replay is the most literal version of that yet: you do the task, Codex watches, and next time it handles it without you.

What Record & Replay actually does

The feature shipped in version 26.616 of the Codex macOS desktop app.

The mechanic: open the Codex app, start a recording, and complete a task yourself — navigating screens, uploading a file, filling in metadata, submitting a form. Codex observes every step. When you're done, it converts that sequence into a Skill: a saved, named workflow it understands well enough to run on its own.

Next time you need that task done, you invoke the Skill by name. Codex runs it autonomously using Computer Use — the same capability that lets it operate your screen the way a human would.

If you've connected Codex to tools like Google Drive, YouTube Studio, or Notion via plugins, it can also call their APIs directly rather than controlling the UI — so a single Skill can pull from a spreadsheet via plugin, switch to YouTube Studio via browser control, and verify via API. This makes skills more reliable than pure screen-control when the interfaces are stable.

Skills can be personal or shared with a team. Save a Skill to a repository directory and anyone on the team can use the same workflow — one person records it, everyone benefits.

What it requires

This feature is not available to everyone, and the constraints are real:

• macOS desktop app only. This is not in the ChatGPT web interface, not on Windows, not on Linux.
• Paid ChatGPT account required. The Codex app is free to download; the Record & Replay feature itself requires a paid tier.
• Computer Use must be enabled. On personal accounts, you turn it on in settings. On Enterprise accounts, an administrator enables it. Without Computer Use, the feature doesn't function.
• Not yet available in the EU, UK, or Switzerland. Computer Use arrived in the EEA on June 16, 2026 — just two days before this feature shipped. Record & Replay's European availability has not been confirmed by OpenAI.

Before expecting to use this, confirm your plan tier, your OS, and where your organization's Computer Use permissions stand.

What's worth recording

The honest answer: repetitive, procedural, fixed-sequence work.

Good candidates:

• Weekly content uploads with the same fields every time — a YouTube video with its title, description, tags, and thumbnail, submitted through the same studio UI
• Recurring report submissions that follow a standard path through internal tools
• Data-entry sequences that pull from one source and populate another on a consistent schedule
• Formatting pipelines where the same steps happen in the same order to the same type of document

Less suited:

• Workflows that vary meaningfully from run to run (different recipients, different formats, different source data structures)
• Tasks that require genuine judgment — reading a dynamic document and deciding what to do based on its content, or handling exception cases
• Any workflow involving regulated or sensitive data before you fully understand what Computer Use is accessing and logging

The specific-and-stable caveat matters more than it might seem. Record & Replay creates a skill from a demonstrated workflow, but that skill is only as robust as the demonstration. If the UI changes, if it encounters an unexpected prompt, or if a step has a decision branch the recording didn't capture, the Skill may fail silently or take an unintended path. Treat new Skills as pilots and check their outputs for the first few runs.

What it doesn't replace

Screen-level control via Computer Use is powerful and fragile in the same places a human operating a computer would be. It works well when interfaces are stable and inputs are predictable. It struggles with CAPTCHAs, unexpected authentication prompts, error dialogs, and UI changes that alter where a button lives.

Record & Replay is a tool for eliminating the genuinely tedious parts of knowledge work — the "I do this the exact same way every Tuesday" category. It is not a replacement for judgment, exception handling, or complex workflows with branching logic. Those still require interactive AI assistance, not a pre-recorded skill.

The broader shift

Record & Replay is one feature in a pattern that's accelerating across all three major AI tool platforms. AI is increasingly designed to observe and repeat, not just respond.

OpenAI is positioning Codex as the layer that coordinates work across multiple applications — not a chat window you type questions into, but an agent that handles procedural workflows end to end. Google is building similar orchestration into Gemini and Google Workspace. Anthropic is expanding what Claude's agentic features can do autonomously.

The direction: the next layer of AI productivity is automation of the repeatable, not just augmentation of the creative. The question worth asking now, before the tools get more capable and more autonomous: which parts of your work follow a fixed script, and what would it mean for your work if those parts ran themselves?

---

Sources

• OpenAI Developers — Codex App Changelog (version 26.616, June 18, 2026): https://developers.openai.com/codex/changelog
• OpenAI Developers — Codex App Automations documentation: https://developers.openai.com/codex/app/automations
• The Decoder, "OpenAI's Codex can now watch you work once and repeat the task forever" (June 20, 2026): https://the-decoder.com/openais-codex-can-now-watch-you-work-once-and-repeat-the-task-forever/