Codex Now Controls Windows Apps: SaaS Take

What did OpenAI ship for Codex on Windows on May 29?

On May 29, 2026, OpenAI brought Computer Use to the Codex app on Windows. Eligible users can now let Codex see the screen, click, and type inside Windows apps. The release also added remote continuation from iOS, Android, and Mac, plus Codex Profiles that show usage and token activity per project.

How does Codex Computer Use actually work?

Codex observes your screen, then takes actions like clicking buttons and typing text to complete a task you describe. It works much like a person driving the machine, but automated. You set the goal, Codex executes the steps, and you watch or step in. It is built on OpenAI's GPT-5.5 model under the hood.

Why does desktop control matter for SaaS teams?

Because it widens what you can automate beyond code and the browser. A coding agent that drives a full Windows desktop can run tools that have no API, handle repetitive setup work, and connect steps across apps. For a lean SaaS team, that shifts the line between what you build, what you buy, and what you automate.

When should a founder let an agent write a full pull request?

Only when a human reviews every line before it merges. Agents are useful for drafting changes and handling boilerplate, but unreviewed generated code is a real risk. Let the agent open the pull request, then treat it exactly like a junior developer's work: read it carefully, test it, and approve it deliberately, never automatically.

Where does the human review gate belong?

At the merge step, always. Let agents draft, explore, and propose freely in a branch, but require human sign-off before anything reaches your main codebase or production. The review gate is the one control you do not automate away. It is what separates fast, safe agent use from shipping code nobody understood.

Which Codex features are new in this release?

The headline addition is Computer Use on Windows, letting Codex operate desktop apps. OpenAI also added remote continuation, so you can start a task on desktop and follow it from your phone, plus Codex Profiles showing usage and token activity. Together they make Codex more of a persistent, observable assistant than a one-off tool.

Should small teams adopt Codex now or wait?

Pilot it on low-risk work now. OpenAI was named a Leader in enterprise coding agents by Gartner on May 27, 2026, and Sam Altman offered new business customers two months free in May, so the cost of testing is low. Start with internal scripts and boilerplate, keep humans reviewing, and expand only as trust builds.

Will agent coding replace junior developers?

No, but it changes the job. Agents handle more boilerplate, so junior developers spend less time on rote code and more on reviewing, testing, and judgment. The skill that matters rises from typing code to evaluating it. Teams still need people who can read generated work critically, which is exactly what juniors should learn now.

Can Codex safely touch a live Webflow project?

Be very careful here. For custom code in a Webflow project, let Codex draft changes in a separate branch or staging environment, never directly on your live site. Review everything, then publish through your normal Webflow workflow yourself. Treat the agent as a drafting tool, and keep the publish action firmly in human hands.

How do I measure ROI on a coding agent?

Track time saved against cost and rework. Note hours the agent removes from routine tasks, subtract its subscription and token cost, and watch how much generated code gets reverted or fixed. If rework stays low and hours saved are real, the agent pays off. If you are constantly cleaning up after it, it does not.

Weighing how much to hand to coding agents? Pair this with my piece on Claude Code's dreaming feature, Simon Willison on Anthropic versus OpenAI, and my Webflow Cloud GitHub deploy tutorial. Let's chat.