Published July 21, 2026
For two years, "AI agent" mostly meant a chatbot with better marketing. That changed this month. A computer-use agent is an AI that operates software the way you do: it looks at the screen, figures out what to click, clicks it, reads the result, and keeps going. No API required, no integration to build. If a human can use the app, so, in principle, can the agent. And within roughly three weeks this summer, Google, OpenAI, and Microsoft each shipped their serious version of one.
The timing is not a coincidence; it is a land grab. Whoever's agent ends up doing your busywork also ends up deciding which apps, which browser, and which cloud that busywork runs through. This piece walks through what each of the three actually shipped, per their own announcements, then gets to the question the launch posts skip: what should you let these things do at work today, and what should stay firmly human?
Google: agents across every screen
Google's June recap, published July 1, put Computer Use in Gemini 3.5 Flash and described the goal plainly: agents that see, reason, and act across desktop, mobile, and browser. The breadth is the story. Google is not confining its agent to one surface; the same capability is meant to work on your laptop, your phone, and inside a browser session.
Shipping computer use in Flash, the fast, cheap tier rather than a premium flagship, is a pricing signal as much as a technical one. Google wants agent actions to be cheap enough that developers sprinkle them everywhere, not a luxury feature you ration. For anyone building automation on top of an API, that makes Gemini the most interesting of the three. For a non-developer at a desk, it is currently the least packaged: the capability is real, but Google is handing you an engine where the other two hand you a car.
OpenAI: the agent moves into the app, the browser moves into the cloud
OpenAI's July 9 was busy. ChatGPT Work launched as an agent aimed squarely at task completion, and the desktop app was unified the same day so that Chat, Work, and Codex live in one place. The pitch is that you hand Work a task, not a conversation, and it goes off and completes it.
The more telling move came alongside it. The same day, OpenAI shut down its Atlas browser and announced a cloud-based remote browser for agents. Read those two together and OpenAI's theory of computer use gets clear: the agent should not be steering the browser on your machine, in your session, with your logged-in cookies. It should do its browsing on OpenAI's infrastructure, remotely, and bring you back the result. That is a genuinely different safety posture from an agent puppeting your own screen. It also means the work happens where you cannot watch it as directly, a tradeoff we come back to below.
Microsoft: agents you can see in the taskbar
Microsoft's contribution sounds boring and is anything but. Per the July 1 Microsoft 365 Copilot release notes, long-running agent tasks are now visible in the Windows taskbar, and Agents support scheduled prompts.
Sit with that first item. An agent task with a taskbar presence is an agent being treated as a first-class citizen of the operating system, something you can glance at, check on, and interrupt, the way you would a file transfer. Of the three vendors, Microsoft is the one designing for the moment an agent runs for forty minutes while you do something else, which is the moment this category becomes real office infrastructure instead of a demo. Scheduled prompts push the same direction: an agent that runs your Monday report every Monday without being asked is closer to a coworker than a chatbot. Less flashy than "see, reason, and act." Possibly more useful this quarter.
At a glance
| Ecosystem | What shipped | Where it acts | Distinctive bet |
|---|---|---|---|
| Google Gemini | Computer Use in Gemini 3.5 Flash | Desktop, mobile, and browser | Cheap, broad capability for builders |
| OpenAI | ChatGPT Work in the unified app (July 9) | Cloud-based remote browser, off your machine | Agent browsing happens on OpenAI's infrastructure |
| Microsoft Copilot | Long-running tasks in the taskbar, scheduled prompts | Windows and Microsoft 365 | Agents as visible, schedulable OS citizens |
What to trust them with today
Now the honest part. Launch posts describe capability. Offices run on reliability, and these are not the same thing. Based on what the vendors themselves have shipped and how they have scoped it, here is our assessment of where the line sits in July 2026.
Reasonable to hand over now:
- Research and gathering: pulling information from multiple sites or documents into one draft, the core ChatGPT Work pitch.
- Recurring report drafts on a schedule, exactly what Copilot's scheduled prompts exist for, provided a human reads the output before it goes anywhere.
- Form-filling and data shuffling between systems that lack integrations, with a review step before submission.
- Long boring multi-step tasks where a wrong step is annoying rather than expensive: reorganizing files, drafting bulk responses, triaging a backlog.
Keep human, for now:
- Anything that spends money or commits the company: purchases, contracts, payroll, vendor payments. An agent that misreads a screen can misclick a confirmation.
- Irreversible actions: deletions, sends to external recipients, production changes. Agents fail confidently, and there is no undo on a sent email.
- Sensitive logged-in sessions: banking, HR systems, admin consoles. OpenAI's remote-browser design exists precisely because letting an agent loose in your own authenticated session is the risky version.
- Judgment calls with people on the other end: performance feedback, client escalations, anything where the cost of a tone-deaf action lands on a relationship.
The pattern in that split is simple: hand over work where errors are cheap and reviewable, keep work where errors are expensive and instant. Capability will keep improving. The economics of an irreversible mistake will not.
Permission hygiene
Whichever agent you adopt, a few practices cost nothing and prevent the stories that end up in incident reports. Run agents in a separate browser profile with its own logins, so a confused agent cannot wander into your personal accounts. Grant access to the specific apps a task needs, not everything, and prefer setups that ask before acting over ones that ask forgiveness. Watch the first ten runs of any new task before you stop watching; Microsoft's taskbar visibility makes this easy, which is partly why we rate it. And write down, somewhere your team can see, which tasks are approved for agents. The alternative is finding out what your coworkers automated when one of their agents does something memorable.
Verdict, by who you are
There is no single winner here, because the three products are barely competing for the same buyer yet. If your company lives in Microsoft 365, Copilot's agents are the obvious start: visible, schedulable, and already inside the tenant your IT department controls. If you are a ChatGPT household or team, ChatGPT Work in the unified app is the natural move, and the remote-browser architecture is a real point in its favor for risky web tasks. If you are a developer or a tinkerer building your own automation, Gemini 3.5 Flash's computer use across desktop, mobile, and browser is the most flexible raw material of the three. And if you are none of these, if you just want less busywork with no setup, wait a quarter. All three vendors shipped their opening move this month, not their finished product, and the second version of each will be judged on the metric none of the launch posts mention: how rarely it needs to be rescued. Our best AI agents roundup covers the wider agent field beyond these three, including the tools that do not need to touch your screen at all.