OpenAI Codex — Record & Replay: Show It Once, Never Do It Again

📅 June 22, 2026

🏷️ 🎬 Product Launch

⏱️ 60-Second Summary

OpenAI shipped Record & Replay on June 18, 2026, as part of Codex app version 26.616. Instead of writing a prompt or building a workflow from scratch, you simply perform a task while Codex watches. The agent then converts that live demonstration into an inspectable, editable SKILL.md — a natural-language description of what the workflow is trying to accomplish. Next time around, you hand Codex the values that are different (a new file, a different date range, an issue number), and it replays the skill autonomously using Computer Use, browser actions, and installed plugins. The feature is available to ChatGPT Plus, Pro, Business, Enterprise, and Edu subscribers, though it is currently limited to macOS and not yet available in the EEA, UK, or Switzerland.

🤔 Why This Matters

Robotic Process Automation (RPA) tools like UiPath and Automation Anywhere have existed for years, but they record pixel coordinates and UI element identifiers — brittle recordings that shatter the moment a button moves or a UI is redesigned. Codex takes a fundamentally different approach: it generates a semantic, natural-language skill rather than a coordinate map. On replay, its language model interprets the skill description against the current screen state, making it semantically adaptive rather than coordinate-literal. This closes a long-standing gap between AI assistants (which understand intent) and automation tools (which execute reliably) — and it does so through a zero-configuration, zero-code user experience.

👨‍💻 Developer Impact

Developers and power users gain the ability to record once and delegate forever. Repetitive operational workflows — filing expense reports, submitting time-off requests, creating GitHub issues, running deployment checklists — can be captured as skills without writing a single line of automation code. Because skills are plain SKILL.md files, they are version-controllable, diff-friendly, and auditable. Developers can also ask Codex to refine a generated skill post-recording, allowing iterative improvement. Engineers building internal tooling can use Record & Replay to rapidly prototype workflows that previously required dedicated RPA infrastructure or custom scripts.

🌍 User Impact

For non-technical users, Record & Replay dramatically lowers the barrier to automation. Rather than learning a workflow builder or scripting language, anyone can demonstrate a task and have Codex replicate it. The Team Sharing capability multiplies this effect: one employee's recorded workflow can become the entire department's automation. A single HR manager recording the company's expense-submission process, for example, instantly gives every team member a ready-to-use skill. The net result is a democratisation of automation — no specialist required.

🔥 Key Features

• Zero-prompt automation: Demonstrate a task; Codex watches and learns — no prompt engineering needed.
• SKILL.md generation: Recordings produce an inspectable, editable natural-language skill file covering purpose, required inputs, step-by-step actions, and result-verification criteria.
• Semantic replay: On replay, Codex interprets the skill against the live screen state using its language model, making it resilient to minor UI changes.
• Parameterised execution: Supply only the values that differ each run (file, date range, issue ID); Codex handles the rest.
• Computer Use + browser actions + plugins: The replay engine combines all available Codex capabilities for full end-to-end automation.
• Team sharing: Skills can be shared across an organisation, turning individual recordings into team-wide assets.
• Available to: ChatGPT Plus, Pro, Business, Enterprise, and Edu subscribers.

⚠️ Current Limitations

• macOS only at launch — Windows and other platforms are not yet supported.
• Not available in the European Economic Area (EEA), United Kingdom, or Switzerland.
• Requires Computer Use to be enabled, which narrows the eligible audience further.
• Skills are only as reliable as the clarity of the original recording — complex branching workflows may require manual refinement.

🔗 Official Source

https://openai.com/codex