Pull to refresh
Logo
Daily Brief
Following
Why Ranks Sign Up
OpenAI extends Codex agent to control Windows desktops

OpenAI extends Codex agent to control Windows desktops

New Capabilities

Computer-use agents move off the browser and onto the operating system, and Microsoft now builds its own

June 2nd, 2026: Microsoft announces Windows Agent Platform, Scout, and Project Polaris at Build 2026

Overview

OpenAI's Codex can now see and click inside Windows applications, a capability Anthropic's Claude has offered since October 2024. Microsoft joined them at Build 2026 in early June, launching its own Windows Agent Framework as open-source software and introducing Scout, an always-on background agent for Microsoft 365.

Microsoft also unveiled Project Polaris, an in-house model to replace OpenAI's GPT-4 in GitHub Copilot by August. Nadella now runs competing agents from OpenAI, Anthropic, and his own team across the same products, while commercial ties to both external labs continue.

Why it matters

If Codex can drive a Windows PC, it can operate your spreadsheets, accounting software, and inbox without you watching.

Questions about this story

0

How does this compare to what Claude has been able to do on Macs?

Claude has been doing Mac computer use since Cowork launched Mac-first in early 2026 — and already reached full Windows parity in February 2026, meaning Codex is replicating capabilities Claude users on both platforms already have.

Why it matters: OpenAI is catching up, not leaping ahead: by the time Codex gains Windows desktop control, Claude had already closed the Mac-to-Windows gap three months earlier.

  • Claude's Cowork feature launched Mac-first, letting Claude see the screen, move the cursor, click buttons, and type into any app — the same core actions Codex now does on Windows.
  • On February 10, 2026, Anthropic shipped Cowork for Windows with what it called complete feature parity; Windows x64 is supported, though ARM devices (Surface Pro X, some Copilot+ PCs) are not.
  • In March 2026, Mac users got an extra layer: remote session handoff, letting you assign tasks from your iPhone and resume on your desktop — a capability not yet part of Codex's Windows release.
  • Claude runs a fallback hierarchy before touching the screen — API connectors first, then browser, then direct system control — and requires explicit per-app permission; Codex's Windows integration is more direct, which is faster but carries more surface area for a wrong click.
Room for disagreement
  • Anthropic frames its permission hierarchy and per-app approval as a safety feature; critics argue it makes Claude slower and more cumbersome for power users who want continuous autonomous operation — the kind of seamless control Codex's direct Windows integration is designed to deliver.
AI-generated with web search — may be wrong. Check the linked sources.

Key Indicators

72%
Windows desktop share
Share of desktop and laptop computers running Windows, per StatCounter.
20 months
Since first computer-use agent
Anthropic launched Claude Computer Use in October 2024. Microsoft announced its own Windows agent platform at Build 2026 in June 2026.
Web → OS → OS-native
Surface area shift
OpenAI's Operator was browser-only. Codex reached the Windows desktop in May 2026. Microsoft's Windows Agent Framework now builds agents directly into the OS layer.
Mobile + Mac
Remote continuation
Users can start a Codex task on one device and resume it on another, raising the share of work an agent can carry alone.

Voices

Curated perspectives — historical figures and your fellow readers.

Ever wondered what historical figures would say about today's headlines?

Sign up to generate historical perspectives on this story.

Play

Exploring all sides of a story is often best achieved with Play.

Log in to play. Track your picks, climb the leaderboards. Log in Sign Up
Predict 5 ways this could play out. Contrarian picks score more — points lock when the scenario resolves. Log in to play
Higher or Lower Two numbers from this story. Guess which is bigger. 5 rounds to set a streak. Log in to play
Timeline Five events from this story — drag them oldest to newest. Log in to play
Connections Sixteen names from the news. Find the four hidden groups of four. Log in to play

People Involved

Organizations Involved

Timeline

October 2024 June 2026

7 events Latest: June 2nd, 2026 · 1 month ago
Tap a bar to jump to that date
  1. Microsoft announces Windows Agent Platform, Scout, and Project Polaris at Build 2026

    Latest Product launch

    Microsoft open-sourced the Windows Agent Framework v1.0 and introduced Scout, its first always-on agent for Microsoft 365, at Build 2026 in San Francisco. The company also unveiled Project Polaris, an in-house AI model to replace OpenAI's GPT-4 in GitHub Copilot by August, with the Windows Agent Runtime slated for general availability in November 2026.

  2. Codex gains Computer Use on Windows

    Capability launch

    OpenAI extends ChatGPT Codex to see, click, and type inside Windows applications. Update also adds remote continuation from mobile and Mac, and new Codex Profiles with token tracking.

  3. Microsoft makes Claude the default AI in Excel and PowerPoint

    Product launch

    Microsoft set Anthropic's Claude Sonnet as the default model in Excel and PowerPoint for most Microsoft 365 Copilot users worldwide. Word is scheduled to follow in summer 2026.

  4. Microsoft and OpenAI end exclusive cloud deal

    Partnership change

    Microsoft and OpenAI restructured their 2019 agreement. OpenAI can now serve customers from competing clouds. Microsoft retains a four-month head start on new model releases and an IP license through 2032, with revenue share payments capped and continuing through 2030.

  5. OpenAI's Chronicle feature draws Recall-style privacy scrutiny

    Privacy concern

    OpenAI's Chronicle captures screen context to build memories for Codex, storing frames locally and sending selected clips to OpenAI servers. Security researchers drew parallels to the Microsoft Recall backlash of May 2024.

  6. OpenAI launches Operator

    Capability launch

    OpenAI's first agent product, Operator, ships inside a sandboxed remote browser. It can fill forms and complete web tasks but cannot reach the desktop.

  7. Anthropic ships Claude computer use

    Capability launch

    Anthropic releases a public beta letting Claude take screenshots, move a cursor, and type into any application. The first general-purpose computer-use model.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

1995-1998

Browser wars (1995-1998)

Netscape Navigator dominated the early web. Microsoft bundled Internet Explorer free with Windows 95, used its OS distribution advantage, and overtook Netscape's share within three years. The Justice Department later sued Microsoft for tying.

Then

Netscape lost most of its share by 1998 and sold to AOL in 1999. Internet Explorer reached over 90% browser share by 2002.

Now

The case established that controlling an operating system gave a vendor decisive leverage over any new layer that ran on top of it. That precedent is now relevant again as agents become a new layer.

Why this matters now

Codex Computer Use is a third-party agent running inside a Microsoft OS. If Microsoft ships its own equivalent inside Windows, the same distribution dynamic that ended Netscape applies again.

May-June 2024

Microsoft Recall backlash (May 2024)

Microsoft announced Recall, a Windows feature that took continuous screenshots of the user's desktop so an AI could search them later. Security researchers showed the local database could be read by malware. Public reaction was sharply negative.

Then

Microsoft delayed the launch, made Recall opt-in, encrypted the database, and moved the screenshots behind Windows Hello authentication.

Now

Set a template for how the public reacts to AI features that watch a user's screen. Vendors now ship such features behind explicit consent and on-device storage.

Why this matters now

Codex on Windows raises a sharper version of the same question. An agent does not just record the screen; it acts on it. Public tolerance for that is still being tested.

October 2024

Anthropic launches Claude computer use (October 2024)

Anthropic released a public beta of Claude that could take screenshots of a user's screen, move a cursor, and type into applications through an API loop. It was the first generally available computer-use model from a frontier lab.

Then

Developers built early agents for QA testing, browser automation, and workflow scripting. Adoption stayed mostly inside engineering teams.

Now

Set the technical template every major lab is now copying, including the screenshot-plus-action loop OpenAI's Codex now uses on Windows.

Why this matters now

OpenAI's Windows launch is the second act of the story Anthropic started 19 months earlier. The race is now about distribution, not capability.

Sources

(14)