The launch of OpenAI's Agent has been the subject of widespread speculation, with rumors swelling since November of the previous yearSources indicated that OpenAI was gearing up for a significant release in January 2025, and more recently, breaking news narrowed it down to the end of this month.

In a groundbreaking move, OpenAI has officially activated the Agent era with the introduction of its first Agent, known as OperatorThe world of technology was abuzz during a live-streamed event in the early hours of January 24, Beijing time, during which Operator took center stage, capturing global attention.

The Operator represents a true digital Agent, able to emulate human manipulation of computers by directly interacting with web pages through clicks, scrolling, and typing, thereby accomplishing a variety of tasksIn essence, Operator functions like a self-aware digital employee capable of navigating websites, filling out forms, ordering goods, and making restaurant reservations, effectively taking on mundane yet complex operations on our behalf.

Prior to this, OpenAI took a preliminary step by launching the "Tasks" feature, which aimed to transition ChatGPT from a passive AI chatbot into an active digital assistant capable of executing tasksThe advent of Operator signifies an important leap for OpenAI, marking a transformation from "passively processing information" to "actively completing tasks,” thereby paving the way for the development of Artificial General Intelligence (AGI).

At this juncture, it is important to highlight that Operator is currently in a research preview phase and is exclusively available to ChatGPT Pro users in the United States, priced at $200 per month, with Plus users left out of this experience

Advertisements

Unlike competitors such as Claude’s Computer Us and ZhiPu’s GLM-PC, which directly operate users' computers, Operator operates in the cloud by launching an online "browser" for performing tasks.

To grasp the significance of Operator, one must consider real-world scenariosDuring OpenAI's live demonstration, observers witnessed how the AI adeptly navigated the digital realm, completing tasks much like an experienced surfer navigating the web.

For instance, in a demonstration, Operator was tasked with reserving a table for two at "Beretta" restaurant for 7 PM that eveningWhile this instruction might involve a few searches and filters for a human, it posed a notable challenge for the AI.

Upon receiving the reservation request, Operator would analyze the requirements and launch a cloud-based browser backend, methodically searching for the restaurant and initiating the reservation processUsers could observe Operator's every click, scroll, and typed input, mirroring human interaction.

The efficiency exhibited by Operator was indeed astonishingIt quickly activated the built-in browser, scanning the contents on the screen and analyzing the webpage structure to locate search fields and various filtersThe entire process unfolded seamlessly, resembling the actions of an actual person managing the booking.

When Operator discovered that there were no reservations available at 7 PM, it proactively searched for alternatives close to the user's request, eventually suggesting "7:45 PM" for the reservation.

Similarly, if the 7:45 PM slot was claimed by another party, Operator adeptly presented two other options: "6:15 PM" and "8:15 PM."

Moreover, during tasks like purchasing groceries, Operator demonstrated its ability to perform continuous operations, such as searching for items and adding them to the shopping cart

Advertisements

Before finalizing the purchase, it prompts the user to reclaim control for confirmation and payment (logging into accounts while retaining authentication), allowing for any last-minute additions or modifications.

Coupled with OpenAI’s earlier introduction of the "Tasks" feature, one could easily envision a future where Operator could routinely replenish household supplies.

From official demonstrations and select user experiences, it's evident that in a variety of contexts such as shopping and ticket booking, Operator has showcased remarkable adaptability and versatility, successfully handling various tasks with ease.

Additionally, as previously stated, users can keep track of every action performed by Operator, or opt out of observation, allowing the agent to execute another task while the user engages in different work until prompted for confirmation by Operator.

Both the official demonstrations and tests from YouTubers highlight these capabilitiesBut how exactly does Operator achieve all of this?

The key to Operator's ability to operate a computer like a human lies in OpenAI's custom-built "Computer Usage Agent" (CUA). CUA is based on GPT-4o’s visual capabilities and advanced reasoning techniques, enabling AI to "understand" and "manipulate" computer interfaces, essentially providing AI with the same ability to interact with Graphical User Interfaces (GUI) that humans possess.

The first step for CUA involves "seeing" and interpreting the screen's contents

Advertisements

It analyzes screenshots to understand various information such as images, text, and recognizes different webpage elements, like buttons, links, and input fieldsThis process mirrors how humans visually interpret the world.

Next, CUA reasons and makes judgments based on the user's instructions and the content it "sees," determining what action to take nextFor example, when tasked with booking a restaurant, CUA deduces it must first visit the restaurant reservation website and input keywords into the search field. This mimics human problem-solving processes.

Simultaneously, CUA executes relevant actions such as moving the mouse, clicking, and typing on the keyboardThese operations are executed with precision, akin to how we control computers using a mouse and keyboardBecause of this generalized interaction capability, Operator doesn’t rely on websites to provide API interfaces, making it viable across virtually any web page.

To enhance the intelligence and fluidity of operations, CUA operates through an iterative loop of "observing," "thinking," and "acting" until the task is completedWhen challenges arise or errors occur, Operator can self-correct using its reasoning skillsIf faced with difficulties or requiring user interaction, Operator can return control to the user.

Additionally, OpenAI has cleverly opted to run the browser in the cloud rather than directly accessing users' computers, which could lead to concerns surrounding "occupancy," "privacy," and "environment."

The first two concerns are relatively straightforward. "Occupancy" refers to the issue where users may be unable to continue other operations while the agent is manipulating the computer, resulting in downtime

As for "privacy," it is self-evident that user computers often house numerous sensitive files and information.

The issue of "environment" relates to the complexity of the typical user's computing environment, which may face various system or software bugs, and even launching software can present permission issues across platforms like Windows, macOS, or Linux.

In contrast, OpenAI appears keen to avoid "taking big steps that might lead to issues," by confining usage scenarios to the most universal application—the "browser"—and using cloud operations to ensure a uniform, private, and background-capable operational environment.

While OpenAI is not the first major model provider to develop true Agents, the integration of these technologies with thoughtful product design not only signifies a leap from "passively processing information" to "actively completing tasks" but also positions Operator as more user-friendly for the mainstream public compared to Claude's Computer Us or ZhiPu's GLM-PC.

The past year has rendered the concept of Agents a common consensus in the AI industryHowever, many offerings marketed as "Agents" merely involve minimal contextual customization, with role-playing Agents executing tasks by breaking down directives without demonstrating true autonomous action.

Essentially, these remain simple software modules rather than genuine autonomous Agents.

On the other hand, true Agents representative of the large model epoch ought to operate as humans do, executing tasks and managing actions such as operating computers and performing duties, thus directly replacing humans in unnecessary operations.

This distinction is critical, helping to differentiate between mere hype and genuine technological breakthroughs while clarifying the value propositions of Claude Computer Us, Honor YOYO Agent, and OpenAI's Operator.

However, it is important to understand that Operator and other similar "true Agents" are still in the early stages of exploration

Advertisements

Advertisements