computer use (CUA)

advancedspreading

An AI capability where the model handles the computer like a human: it looks at the screen, clicks, types, and scrolls.

In plain terms

Traditional chat AI only provides answers: you ask, it answers, and the clicking remains on you. Computer use reverses this: the AI sits at the computer itself, opens the program, fills out the form, and copies data from one window to another.

It is as if you asked a friend for remote assistance, who takes over your mouse and keyboard. They see the same thing you do and press the same buttons, except here an AI model sits in the friend's place.

It is important because most programs lack a dedicated machine interface (API), offering only a human user interface. If an agent can work through the screen, it can theoretically use anything you can: legacy corporate systems, online stores, spreadsheets.

It differs from traditional automation because it does not execute a pre-written sequence of steps. Before each step, it checks what is currently on the screen and adapts accordingly, making it more flexible but slower and more error-prone.

How it works

You provide the task in everyday language, for example: 'download last month's invoices and enter the amounts into the spreadsheet.'
The agent receives a screenshot of the machine (or of a separate, isolated virtual machine), and the model 'looks' at what it shows: buttons, fields, menus.
The model decides the next step and issues it as an instruction: move the mouse to this point, click, type this text, scroll down.
A program executes the step, a new screenshot is taken, and the loop restarts. This repeats until the task is completed or the agent gets stuck.
At sensitive points (payment, login, deletion, messaging), a well-configured system pauses and requests your approval.

An everyday example

You ask the AI assistant to find three accommodations near Lake Balaton for the specified weekend and compare the prices. In front of your eyes, it opens a browser, enters the dates, clicks through the results, and compiles the findings into a table; it stops before booking and asks you.

Who is affected

employeeRepetitive, click-heavy office work (data transfer, form filling, report downloading) can increasingly be handed over to an agent. → Work & HR

entrepreneur, executiveIt can automate even legacy systems that have no API, but you must decide what access to grant it.

developerIt is a new tool for testing and automation, but you must build an isolated environment, access control, and defense around it. → Code

everyoneIf an AI clicks on your computer in your name, it gains access to your logged-in accounts as well, so you must know what you allow it to do.

A common misconception

MythComputer use means that AI now handles the computer as quickly and reliably as a human.

In realityModels have improved significantly, but agents typically still progress more slowly than an experienced user and may get lost or make mistakes on complex, multi-step tasks. Without supervision, they are currently better suited for simple, well-defined tasks.

Limits and risks

It can be slow and expensive: every step requires processing a new screenshot, so a human task taking a few minutes may take much longer.
It is vulnerable to hidden instructions: text placed on a website or in an email can compel the agent to do something you did not request (prompt injection).
It works with your accounts and permissions, so a mistaken click can cause real damage: a sent email, a deleted file, or a placed order.
Do not trust it with passwords, bank data, or personal documents without verification, and if possible, run it in an isolated browser or virtual machine.

Similar, but not the same

robotic process automation (RPA) – Traditional RPA plays back a pre-recorded, rigid sequence of steps, while computer use examines the screen at every step and decides on the continuation itself.
tool use – Technically, computer use is a type of tool use. In standard tool use, the agent calls a service directly via a programming interface (API), whereas in computer use, it clicks through the user interface designed for humans.
browser agent – The browser agent works only on websites, while computer use can extend to the entire machine, including desktop applications and files.

Related concepts

Broader

AI-agent, tool use, automation

Narrower

browser agent, screenshot interpretation

On the edge

robotic process automation, prompt injection, multimodal model, sandbox, human approval

A brief history

The term was popularized by Anthropic when it released the computer use capability of its Claude model in public beta in October 2024. In January 2025, OpenAI also presented its own solution, called Operator; it named the underlying model Computer-Using Agent, which is where the CUA acronym comes from. Operator was folded into ChatGPT's agent feature in the summer of 2025. Google and other major developers have since released similar capabilities.

Where things stand

As of late 2026, it is available in the assistants and browsers of the major AI developers, and companies are experimenting with it in office workflows. Because of reliability and security concerns, most organizations still use it with human supervision and for limited tasks.

Status: 2026 November

In the news

Appears in 3 Agent Age stories; the curve shows weekly occurrence in the news magazine.

OpenAI releases GPT-6.1 Terra, which it says is strong at computer useOct 26, 2026 · Agent Age
GitHub Copilot can now work with desktop apps through computer useOct 1, 2026 · Code
OpenAI's Agents API offers Codex harness as a service with computer use; data stored only in USSep 30, 2026 · Code

Also known as

computer use · computer use · computer-using agent · computer-using agent · computer-using agent · CUA · computer use agent · AI machine operator

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.