OpenAI launches an AI agent that can perform tasks autonomously

OpenAI launched the preview mode of Operator, an AI agent that can use its browser to perform multiple tasks for you. 

It is currently available to US Pro users ($200 subscription plan), but the new feature will soon be rolled out to Plus, Team, and Enterprise tiers. 

OpenAI details how it works and its capabilities. The company announced that the Computer-Using Agent (CUA) can perform various tasks on behalf of the users. These include buying groceries, exporting images, downloading files, combining PDFs, and analyzing spreadsheets. 

How Operator Works

Powered by a Computer-Using Agent (CUA), a new AI model, the Operator combines the vision capabilities of GPT-4o with advanced reasoning through reinforcement learning. This new AI model is trained to interact with graphic user interfaces (GUIs) and front-end websites; thus, it can tap into different services without using developer-facing APIs or custom API integrations.

It can see through screenshots and interact using the actions that your keyboard and mouse allow, including navigating menus, using buttons, and filling out web pages. 

What if it makes mistakes? The AI model will use reasoning to correct itself. If it needs assistance or gets stuck, like when it runs into a password field, CAPTCHA check, or a complex interface, it’ll ask the user to take control for smooth collaboration. 

Automating tasks is one of OpenAI’s promises. It can also make restaurant reservations, book travel accommodations, and shop online. Users can choose from several categories in its interface (shopping, delivery, dining, and travel), each offering different automation. 

By activating Operator, a new small window will display a dedicated web browser (a small pop-up window), which the AI agent will use to complete tasks, including explaining its actions. 

Because the AI agent uses its dedicated browser, users can still control what they do on their computer screen while the agent is working. 

Safety Approach

Safety is still a huge concern for Operator, but OpenAI says it’s taken risks into account. The AI agent mitigates misuse by blocking illegal and harmful tasks and not being able to access blacklisted sites, such as adult entertainment and gambling sites, as well as gun or drug retailers.

Since it can also make mistakes, which can be costly, without human supervision, it’ll ask for user confirmation before sending an email or submitting an order, etc, allowing them to double-check everything before the AI agent finalizes a task.

For security reasons, the AI agent also limits the number of open conversations and simultaneous tasks a user can perform at any given time (these limits may change).

Limitations

In a support document, OpenAI says that Operator can’t effectively handle specialized or complex tasks, like managing intricate calendar systems, interacting with non-standard or highly customized web interfaces, or creating detailed slideshows. 

It’ll also refuse to send emails, conduct financial transactions, delete a calendar event, and perform specific higher-stakes tasks.

According to OpenAI, its capabilities will expand soon, but these current limitations will help ensure its operation’s safety, reliability, and accuracy. 

An Agentic Future: The Next Big Thing in Artificial Intelligence

“We believe that, in 2025, we may see the first AI agents “join the workforce” and materially change the output of companies.”  - Sam Altman in a blog post.

AI agents are the new artificial intelligence technologies pitched to change how we use our computers and the web. Instead of just delivering and processing data or information, these AI agents can act on behalf of humans. With Operator’s release, it will soon be clearer for us to see how realistic the vision of AI agents is. 

Check out more AI news here!