OpenAI Codex Record & Replay: show a task once, the agent repeats it infinitely — the end of manual scripting
🔎 Why Record & Replay changes everything for automation
Computer task automation has been a nightmare for decades. Excel macros are fragile. Python scripts require technical skills. RPA tools cost thousands of euros and break as soon as an interface changes.
OpenAI just released a feature that makes all of this obsolete. Record & Replay, integrated into the Codex app version 26.616 on June 18, 2026, allows you to demonstrate a workflow just once on your Mac so the AI agent repeats it indefinitely. Zero lines of code. Zero complex prompt engineering.
This feature is part of OpenAI's wave of aggressive automation: on the same day, Scheduled Tasks arrivait dans ChatGPT, with a dedicated sidebar to manage recurring tasks. But Record & Replay goes further: it's the realization of the "show, don't tell" promise of agentic AI.
The essentials
- Record & Replay captures your clicks, keystrokes, and window changes to create a reusable "skill" in Codex.
- Available on macOS only at launch (version 26.616), for paid ChatGPT accounts. The EEA, the UK, and Switzerland are excluded at launch.
- Works thanks to Computer Use — Codex "sees" your screen and understands the context, not just the click coordinates.
- Skills are editable: you can modify the sequence, add conditions, or adjust the steps after recording.
- Concrete use cases validated by OpenAI: expense reports, leave requests, recurring emails.
Recommended tools
| Tool | Main usage | Price (June 2026, check on openai.com) | Ideal for |
|---|---|---|---|
| Codex (OpenAI) | Automation by demonstration | Included in paid ChatGPT subscriptions | Knowledge workers, office automation |
| Claude Computer Use | AI computer control | Included in Claude Pro/Team/Enterprise | Workflows requiring complex reasoning |
| Macros classiques / Automator | macOS system automation | Free | Simple system tasks, no contextual understanding |
How Record & Replay works technically
Record & Replay relies on Codex's Computer Use layer. It is not a simple macro recorder that memorizes click coordinates.
The agent simultaneously captures three things: your actions (clicks, keystrokes, shortcuts), the visible content of windows (text, buttons, menus), and the logical sequence between steps. This distinction is crucial. A classic macro that clicks at coordinates (x:340, y:520) breaks if the window is moved. Codex, on the other hand, understands that it needs to click the "Submit" button because it recognizes the interface content.
Recording is triggered in three ways according to the official documentation: via the Codex app menu bar, via a visual overlay, or by voice command. Once the workflow is demonstrated, Codex saves it as an editable "skill".
It is this combination — visual perception + contextual understanding + editability — that makes Record & Replay a generational leap compared to previous automation tools. This evolution follows in the footsteps of the arrival of Codex Computer Use on Windows, which had already laid the foundations for AI agent computer control.
The role of GPT-5.3 Codex in the loop
The underlying model is GPT-5.3 Codex, which scores 80 on agentic benchmarks (June 2025). It is not OpenAI's most powerful model — GPT-5.5 dominates at 98.2 — but it is optimized for computer control tasks and sequential workflow execution. The choice is logical: for desktop automation, execution reliability takes precedence over abstract reasoning.
Concrete use cases that change the game
OpenAI validated three scenarios in its official video demo and in its article on AI/TLDR. Each one deserves to be dissected because they reveal what Record & Replay can and cannot do.
Expense reports
This is the flagship case of the demo. The user opens their management tool, navigates to the expense report form, fills in the fields (amount, category, receipt), and submits. The whole process is demonstrated once. Then, Codex can reproduce this workflow with new data.
The benefit is not just speed. It's resilience: if the form changes slightly (a renamed field, a moved button), Computer Use adapts because it understands the semantic structure of the page, not just its geometry.
Leave requests
Same principle: navigating the HR portal, selecting the dates, filling in the reason, submitting. The workflow is captured in 30 seconds and becomes a skill available in one click.
This case clearly illustrates the difference with macros: an HR portal changes frequently (updates, new mandatory fields). A classic macro requires constant maintenance. A Codex skill, thanks to its visual understanding, tolerates these variations better.
Recurring emails
Replying to the same types of emails with the same format. The user shows how they open the message, draft the reply, choose the recipients, and send. Codex reproduces the pattern while adapting to the specific content of each new email.
This is where the limit clearly appears: if the email requires nuanced judgment, Record & Replay is not enough. It reproduces the pattern, not the reasoning. For template responses, it's perfect. For contract negotiations, it's insufficient.
Record & Replay vs automation alternatives
The desktop automation market is crowded. Record & Replay arrives with a clear positioning: zero code, zero configuration, contextual understanding. But how does it actually compare?
Against classic macros (Automator, AutoHotkey, Excel VBA)
Macros are free, built-in, and have been working for 30 years. But they are fragile by nature: they reproduce mechanical sequences without understanding the context. Move a window, change a screen resolution, update a software — the macro breaks.
Record & Replay is fundamentally different because Codex "reads" the screen. It doesn't click at fixed coordinates, it identifies UI elements by their content and role. It's the difference between a blind person memorizing a path and someone who sees where they are going.
Against Anthropic Computer Use (Claude)
Anthropic launched Computer Use with Claude in late 2024, and it has become a mature feature. The key difference: Claude Computer Use works via text instructions ("go to this site, click this button, fill out this form"), not by demonstration.
Anthropic's approach is more flexible for one-off tasks and requires prompt engineering. OpenAI's approach is more accessible for repetitive tasks — you show, you don't describe. Both have their place. Claude with Sonnet 4.6 (score 81.4) excels on workflows that require reasoning along the way. Record & Replay excels at the faithful repetition of a demonstrated pattern.
Against RPA tools (UiPath, Automation Anywhere)
Professional RPA tools are powerful but have three dealbreakers for individuals and small teams: the cost (often several thousands per month), the implementation complexity (weeks of development), and the fragility in the face of UI changes.
Record & Replay directly targets the segment of simple to medium repetitive tasks — exactly where RPA is oversized and underperforming. It's the same phenomenon as when no-code tools ate the bottom of the software development market.
Against automated screen capture
There are tools that capture screen sequences to replay them. But without an AI layer, these tools are even more fragile than macros: they compare pixels, not meaning. If a button changes color or if a font is updated, the image comparison fails.
Computer Use adds the semantic layer that these tools lack. This is what makes Record & Replay viable in production, where raw screen capture is not.
Current availability and limitations
All the power of Record & Replay should not obscure its real constraints. OpenAI launched this feature in a targeted way, and you need to understand exactly what is included and excluded.
Geographic restrictions
The official documentation is clear: Record & Replay excludes the European Economic Area, the United Kingdom, and Switzerland at launch. This is a classic regulatory restriction for computer control features, linked to the GDPR and local legislation on computer monitoring.
A rollout in Europe is planned but undated. For French-speaking users in Europe, this means waiting. For those located outside these zones (Canada, United States, Asia, etc.), the feature is immediately accessible.
macOS only
No Windows version at launch. This is a strategic choice consistent with the progressive deployment of Computer Use — Codex Computer Use arrived on Windows separately, and Record & Replay likely follows the same path with a delay.
The macOS restriction also allows OpenAI to control the execution environment. macOS offers more predictable screen capture and interface control APIs than the Windows ecosystem, which is particularly fragmented.
Paid accounts only
Record & Replay is not available on free ChatGPT accounts. A paid subscription (Plus, Team, or Enterprise) is required. The exact price depends on the plan (check on openai.com, June 2026), but this clearly positions the feature as a professional tool.
Inherent technical limitations
Record & Replay doesn't solve everything. First point: it does not handle multi-day workflows. If your task requires waiting for an email validation and then resuming the next day, Record & Replay alone is not enough — you would need to combine it with Scheduled Tasks.
Second point: applications with anti-bot securities (CAPTCHA, automation detection) can block execution. Computer Use is designed to resemble human interaction, but it is not infallible.
Third point: the quality of the skill depends directly on the quality of the demonstration. If you make a mistake during the recording, the skill will reproduce the error. Fortunately, skills are editable after the fact — this is an essential safety net.
Implications for knowledge workers
Record & Replay is not a gadget. It is a tool that redefines what "office productivity" means. The implications are concrete and measurable.
The time saved is real
According to MyKreaTool, the identified use cases (expense reports, time-off requests, recurring emails) typically represent 2 to 5 hours per week for a knowledge worker. This is purely mechanical time, with no added value, consisting of copy-pasting data from one system to another.
Multiplied by the number of employees in a company, Record & Replay can represent hundreds of hours saved per month. Not by optimizing the process, but by eliminating it entirely.
The decentralization of automation
Until now, automating an office task required three intermediaries: the employee who identifies the need, the developer or RPA consultant who builds the solution, and the IT department that validates the deployment. This cycle can take weeks.
With Record & Replay, the employee demonstrates the task on Monday, the skill is operational on Monday. Automation goes from being an IT project to an individual gesture. It is an organizational paradigm shift, not just a technical one.
The risk of "AI copy-paste" drift
The obvious risk is that employees will automate tasks that shouldn't be automated. An expense report with variable amounts requires human verification. A recurring email may contain nuances that the skill does not capture.
Record & Replay does not replace judgment. It eliminates the mechanics. The boundary between the two is sometimes blurred, and that is where errors will occur. Companies will need to define safeguards — likely an approved list of workflows for automation.
Record & Replay in the broader OpenAI ecosystem
This feature doesn't come out of nowhere. It is part of a coherent OpenAI strategy to transform Codex from a simple coding tool into a general-purpose automation agent.
Convergence with Scheduled Tasks
The simultaneous launch of Scheduled Tasks in ChatGPT and Record & Replay in Codex is no coincidence. The two features complement each other: Record & Replay defines what the agent does, Scheduled Tasks defines when it does it.
Imagine this: you record an email processing skill, then schedule it to run every morning at 8 AM. Automation becomes an intelligent cron job, without a single line of code.
The link with on-premise deployment
The arrival of Codex on-premise in partnership with Dell opens up an interesting perspective for Record & Replay. Companies that hesitate to send screenshots of their work environment to OpenAI's servers may one day be able to use Record & Replay locally.
This is particularly relevant for regulated sectors (banking, healthcare) where on-screen data contains sensitive information. On-premise could be the enabler that transforms Record & Replay from a consumer tool into an enterprise tool.
The dynamic with Microsoft
The context of Microsoft Build 2026 and Project Polaris is also relevant. Microsoft is developing its own Windows Agent Framework, which could eventually compete with OpenAI's approach. But for now, Record & Replay has a significant lead in terms of ease of use.
OpenAI and Microsoft remain strategic partners, but their visions for the AI agent on the computer are gradually diverging. OpenAI is betting on direct demonstration. Microsoft is betting on a system-integrated agent framework. The market will decide.
❌ Common mistakes
Mistake 1: Confusing Record & Replay with a video macro
A video macro records pixels and coordinates. Record & Replay records intents and context. If you treat Record & Replay like a macro, you will be disappointed by the replay limitations on modified interfaces. Treat it like an agent that has "learned" your workflow, and you will unlock its true value.
Mistake 2: Recording a too complex workflow all at once
The longer a skill is, the more fragile it is. If you were to record a 25-step workflow crossing 5 applications, the probability of failure upon replay increases exponentially. Break your workflows down into short skills and chain them together. A reliable 5-step skill is better than a finicky 20-step skill.
Mistake 3: Ignoring the editing step
OpenAI insists that skills are editable. Many users will record an imperfect workflow, see that it fails on replay, and conclude that the feature doesn't work. Best practice: record quickly, test, edit the skill to correct demonstration errors, then try again. Iteration is key.
Mistake 4: Automating high-variance tasks
Record & Replay excels at repetitive tasks with a stable structure. If the content changes radically with each execution (an email whose format varies hugely, a form with unpredictable conditional fields), the skill will fail. Stick to stable patterns: that's where the feature shines.
❓ Frequently Asked Questions
Does Record & Replay work on Windows?
No, only macOS at launch (June 2026). OpenAI deployed Computer Use on Windows separately, and Record & Replay will likely follow with a delay. No date has been announced.
Are my screen data sent to OpenAI's servers?
Yes, at launch. Computer Use requires the agent to "see" your screen, which involves sending screenshots to OpenAI's servers. For sensitive environments, on-premise deployment with Dell could be a future solution, but it is not yet confirmed for Record & Replay.
What is the difference with ChatGPT's Scheduled Tasks?
Record & Replay defines how the agent acts (the workflow). Scheduled Tasks define when it acts (the scheduling). These are two complementary features, not competing ones. You can create a skill with Record & Replay, then schedule it with Scheduled Tasks.
Can a skill be shared with colleagues?
OpenAI has not detailed skill sharing at launch. As it stands, skills seem tied to the user account. Team sharing would be a logical addition for ChatGPT Team and Enterprise accounts, but we need to wait for an update.
Does Record & Replay replace autonomous AI agents like AutoGPT?
No. Autonomous AI agents operate in a self-directed manner with broad goals. Record & Replay is deterministic: it faithfully reproduces a demonstrated workflow. These are tools for different needs — autonomy vs. reproduction fidelity.
✅ Conclusion
Record & Replay is the most accessible automation feature ever launched by a major AI player. No code, no prompt, no configuration: you show, Codex repeats. Knowledge workers who spend hours on repetitive tasks finally have a tool that lives up to the "show, don't tell" promise. The macOS restriction and geographic exclusion temper the enthusiasm, but the direction is set. Manual scripting has just shown its age. To delve deeper into the meilleurs agents IA autonomes and understand where Record & Replay fits into the agentic ecosystem, our comprehensive guide covers all the key players in the market.