OpenAI is reportedly preparing to add a new kind of capability to ChatGPT: a browser agent called "Operator" that can act inside a web browser on a user’s behalf. According to a report from The Information, the feature is planned as a new ChatGPT feature for browser control later this week.
The reported launch would mark a practical step toward AI agents that do more than answer questions. Instead of only producing text, Operator is described as a system that can open a browser view, carry out actions, ask for missing details, and let the user step in when needed.
What Operator is expected to do
Operator is reportedly designed around browser-based tasks. The feature will offer several task categories, including food and events, delivery services, shopping, and travel planning.
Each category is expected to include suggested prompts, giving users a starting point rather than forcing them to write a detailed instruction from scratch. That matters because browser agents are not only about what the model can do, but also about how clearly a user can ask for the right outcome.
When a user enters a prompt, a mini-screen opens inside the chatbot. That mini-screen shows a browser window and displays the agent’s actions in real time.
The system is also expected to ask follow-up questions when the task requires more information. The source gives the example of a restaurant reservation, where the agent may need details such as the time and number of guests.
Why browser control changes the ChatGPT experience
A browser agent changes the role of ChatGPT from a conversational assistant into something closer to a task assistant. In the reported design, the user gives an instruction, then watches the agent work through a browser session.
That distinction is important. A normal chatbot can explain how to book travel, compare shopping options, or plan an event. A browser control feature aims to operate within the actual web workflow connected to those tasks.
The source describes several features that keep the user involved:
- Users can take control of the screen while Operator is working.
- Users can save operator tasks.
- Users can share operator tasks with other users.
- The system can work with websites that require users to log in.
There is one reported exception: Google’s Gmail is reportedly not included in the login-capable workflow. The source does not provide further detail on why Gmail is treated differently.
This combination of automation and visibility suggests a controlled model of browser assistance. The agent acts, but the user can see what is happening and intervene when necessary.
The launch follows earlier reports and delays
The latest report builds on earlier coverage. On November 14, 2024, Bloomberg reported that OpenAI was planning an AI assistant called "Operator" for a January launch, citing two people familiar with the matter.
According to that report, OpenAI executives announced in an internal meeting that the tool would first launch as a research preview and through an API for developers. The source describes Operator as a general-purpose assistant with a focus on browser-based tasks.
On January 7, 2025, The Information reported that OpenAI might release Operator this month, confirming earlier Bloomberg coverage from November. That report said the launch had apparently been delayed by safety concerns around "prompt injections".
The source defines prompt injections as a security vulnerability where users can manipulate an AI system into ignoring its built-in rules and restrictions. The article also says this has been a known issue since at least GPT-3, and that there is still no reliable defense against prompt injection attacks.
The concern is especially relevant for autonomous AI agents. When an agent reads and acts on web content, users have less direct oversight of what content the model processes. That creates a different risk profile from a chatbot that mostly responds to a user’s own text.
How Operator fits into the AI agent race
Operator is part of a broader industry push toward AI agents. Several major AI labs are developing assistants that can automate multi-step tasks with minimal user supervision.
The source notes that Anthropic has already launched an assistant that processes screen content and performs real-time actions. Microsoft has integrated automation features into its Copilot platform.
Google is developing "Project Jarvis," described as a Chrome-based AI assistant designed to handle tasks like online shopping and travel booking. The company plans to launch it alongside its new Gemini language model in December.
OpenAI CEO Sam Altman views AI agents as the next phase of AI growth. The source connects that shift to slower progress in traditional language model development and says Altman suggests the future lies in using existing models more effectively.
There is still no standard definition for agentic AI systems. The source describes them as small programs or prompts that handle individual subtasks and coordinate with other assistants, either within one language model or by bridging different language models and AI systems.
From chatbots to coordinated workflows
The larger goal is workflow automation. By connecting multiple assistants that reliably perform specific tasks, companies aim to automate entire workflows through coordination.
OpenAI has already taken a first public step in this direction by releasing "Project Swarm" on GitHub. The source describes it as an experimental open-source framework that lets developers create and manage multiple-assistant systems.
Project Swarm demonstrates how assistants can transfer control between each other and execute defined task steps with specific tools. According to OpenAI, Project Swarm serves as a practical demonstration of how their assistant concept works in real-world applications.
If Operator launches as reported, it would bring that agent idea into ChatGPT through a browser-focused interface. The immediate use cases may sound familiar, such as shopping, travel planning, delivery services, and food and events. The bigger shift is that ChatGPT would not only describe the next step, but begin performing it inside a visible browser session.