A new kind of AI is growing very fast. It is called a computer-use agent. Unlike a normal chatbot, a computer-use agent can look at a screen and act inside software by clicking, typing, and scrolling. OpenAI says its Computer-Using Agent, or CUA, was built to use graphical interfaces the way people do. (openai.com)
This is already moving from research to real products. OpenAI first launched Operator in January 2025, then folded those browser skills into ChatGPT agent on July 17, 2025. OpenAI’s help center says ChatGPT agent can do online tasks like filling out forms and editing spreadsheets, while the user stays in control. (openai.com)
Other big companies are moving in the same direction. Anthropic offers a computer use tool in its API, with screenshots plus mouse and keyboard control for desktop tasks. Its docs also warn developers to use sandboxed systems, avoid sensitive data, and ask for human confirmation before actions with real-world consequences. (docs.anthropic.com)
Google is pushing this idea too. Project Mariner, Google’s research prototype, is now available in the U.S. for Google AI Ultra subscribers, and Google says it can handle multiple browser tasks at the same time. Google has also said Mariner asks for final confirmation before sensitive actions like purchases. (deepmind.google)
And in Microsoft’s world, this shift became even clearer in 2026. Microsoft says Browse with Copilot is rolling out in the U.S., and on April 22, 2026, it made Copilot’s agentic actions in Word, Excel, and PowerPoint generally available. (support.microsoft.com)
So, the day when AI moves through your computer for you is no longer a science-fiction idea. But it is not full freedom yet. These systems are improving fast, but companies still warn about mistakes, prompt injection, logins, payments, and private data. For now, the smartest way to see a computer-use agent is simple: not as your replacement, but as a helper for the boring clicks. (openai.com)










