OpenAI shipped Codex Computer Use on Windows on May 29. Version 26.527. The agent sees the screen, clicks, and types inside foreground desktop apps. You steer it from the ChatGPT mobile app. EU, UK, and Switzerland are locked out at launch.
Codex is no longer just a CLI. It now drives Windows desktop apps with its own pointer and keyboard.
The agent runs in the foreground. It moves the cursor, types in dialog boxes, debugs GUI tools, takes screenshots back to you. The mobile-remote piece is the kicker: kick off a session from ChatGPT on your phone while your laptop sits on your desk. EU users get blocked at launch on regulatory grounds, which is itself the story.
For PMs building desktop apps: assume your UI will be tested and used by an agent within 12 months. For execs: the procurement bake-off between Codex, Claude Code, and Cursor just added a desktop-control dimension. For Microsoft: OpenAI shipped Computer Use on Windows the same week Copilot moved to token billing. That's not coincidence.
⚡ Why this matters
- First major lab to put a general-purpose desktop agent on the OS most enterprises actually run, not just macOS.
- Mobile-remote steering removes the last excuse for an agent needing constant attention. Start it on your phone in a meeting, check on it after.
- EU/UK/Switzerland geo-block tells you regulators got a closer look and OpenAI flinched.
🔍 What happened
- May 29, 2026. Codex changelog entry: "Computer use and mobile access on Windows." Version 26.527.
- Agent operates the active Windows session: clicks, types, takes screenshots, debugs GUI apps in the foreground.
- Mobile control: ChatGPT app on iOS/Android can start new threads, send follow-ups, approve actions, review diffs and test results, see screenshots and terminal output.
- Limitation: runs in foreground only. The agent moves the pointer; you can't keep using the same Windows session at the same time.
- Locked out at launch: European Economic Area, United Kingdom, Switzerland.
- Macros, multi-monitor handling, and background mode flagged as upcoming.
💬 Smart takes
- OpenAI changelog (May 29): Codex can now "test apps, debug flows, and review work where your project context lives" on Windows.
- Thurrott (May 29): Codex is the first cross-platform desktop agent from OpenAI that explicitly targets enterprise Windows.
- Windows Forum: mobile remote control turns the agent into a true background worker for the first time. "Steerable from your phone" is the operative phrase.
- Skeptic: running in foreground only means the agent owns your machine while it works. No background mode is a real limit for parallel agent setups.
🧭 Where this goes
- Claude Code ships its own desktop-control feature for Windows within 90 days. Anthropic can't let OpenAI own this surface.
- Cursor responds with a closer integration of OS automation tools (AppleScript on macOS, COM on Windows).
- EU AI Act enforcement on desktop agents becomes the next compliance fight. Geo-blocking is the new normal until classification settles.
- QA and testing teams begin retiring scripted Selenium/Playwright suites for prompt-driven test flows.
- Enterprise IT writes new policies on "agent-owned sessions" by Q4. Permission scopes get formal.
🎯 Implication
- For PMs: design your UI as if a model will read screenshots and operate it. Accessibility labels are now product feature, not afterthought.
- For execs: agentic desktop control is the procurement question this quarter. Get a pilot started with Codex + Claude Code in parallel.
- For security teams: draft an "agent-on-endpoint" policy. Permission profiles, logging, kill switches.