Today. Company is unveiling ChatGPT Agentwhich allows its AI chatbots to browse the web autonomously, conduct extensive research and download and create new documents for its human users by using its own virtual computers.
Are you back? ChatGPT gets its own computer? It can log into the accounts of human users and download or send things for them using that PC? OpenAI says
It’s true, at least in the virtual sense. As the company explains
AI Impact Series Returns To San Francisco – 5 August
Are you ready for the next phase of AI? Join leaders from Block GSK and SAP to get an exclusive look at the ways autonomous agents are reshaping workflows in enterprise – from end-to-end automated workflows to real-time decision making.
Reserve your seat now as space is limited. https://bit.ly/3GuuPLF
“The model can choose to open a page using the text browser or visual browser, download a file from the web, manipulate it by running a command in the terminal, and then view the output back in the visual browser. The model adapts to perform tasks with speed, accuracy and efficiency.”
How to use ChatGPT Agent
The user can engage the agent simply by clicking the ‘Tools” button in the ChatGPT prompt box, opening the menu and selecting the ‘agent mode from the available options’.
Once it’s on, describe a task using plain language and the agent will carry it out in web and local apps environments, combining reasoning and actions that a human could only perform manually on their own computer.
The ChatGPT agent is able to connect to apps such as your personal or professional Gmail, and GitHub so that it can pull useful information – emails or code – from your accounts in order to assist with tasks you request. It can also connect to third-party APIs to pull information, and use connected apps and services.
When a website requires you to log in, a special browser window allows you to do so securely. This lets the agent dig further and perform more personalized tasks like checking your email or filling out forms for you.
Offline, where Operator couldn’t go
ChatGPT builds on and expands the “Operator”agent OpenAI released in Jan 2025. This agent allowed ChatGPT browse the web, fill out forms, make orders, and perform other web-based tasks using a “headless web browser” that OpenAI maintained and offered to each Operator session. Operator, however, was limited to only interacting with web-based applications and websites. It did not include programs that could be run locally on PCs, such as spreadsheet tabulators or slide deck presentation software.
ChatGPT can now browse websites, interact online forms, run codes, analyze data and deliver finished outputs – such as editable spreadsheets or presentations – based on user instructions.
The unveiling follows a report published a few days ago by an independent subscription tech industry site The Information ( ) suggesting that OpenAI will upgrade ChatGPT so as to be a direct competitor to Microsoft’s Office Software Applications (e.g. Excel, Word, PowerPoint, etc.)
Merging Operator and Deep Research to one agent
OpenAI positions ChatGPT as a merger of two of its previous agents — Operator, and Deep Research. The latter was introduced in February 2025 and exhaustively searches the internet through its headless text-only web browser in order to find and compile the information into long and in-depth (hence the title) reports. OpenAI writes about this in a blog:
Operator couldn’t do detailed analysis or write detailed report, and Deep Research couldn’t interact on websites to refine the results or access content that required user authentication. We found that many of the queries users tried with Operator were better suited to deep research. So we combined the two.
The agent can seamlessly switch between a visual web browser and a terminal for running Python code within the same session. It supports a wide range of use cases from analyzing competition and generating reports, to planning trips, summarizing email, or booking appointment.
Users are able to interrupt, redirect or pause any task at any time. The agent will pick up where the user left off.
Access and availability
As of today, subscribers who pay $200 per month for the “Pro” tier have full access to ChatGPT Agent, with a monthly message quota.
ChatGPT plus ($20 per monthly) and Team ($30 monthly) will have access to ChatGPT agent over the next few weeks, with a monthly quota of 40 messages. Credit-based options are available for additional usage.
OpenAI stated in a press release shared with VentureBeat, under embargo, that its ChatGPT Education and Enterprise subscribers will have access to this feature within the next few weeks.
The feature is currently not available in Europe or Switzerland. This will no doubt disappoint residents.
OpenAI was built with safety and control in mind
Now that the agent is able to take actions for users, including on websites logged in or apps connected, OpenAI has implemented extensive safety measures.
These measures include user confirmations prior to taking action, active monitoring for sensitive tasks, as well as technical safeguards that limit unintended behaviors.
The key protections include the following:
- Confirmation Prompts prior to actions such as submitting forms or emails
- Watch Modewhich pauses the execution when a person becomes inactive.
- Resisting high-risk tasks, including financial transfers or privacy violations.
- There is no memory retention during agent sessions.
Classification of domains at high risk
OpenAI, in accordance with its Preparedness Framework treats ChatGPT as a Highly capableagent in the biological domains and chemicals.
Despite the fact that there is no direct evidence to suggest misuse, OpenAI is activating its most stringent safety measures out of caution. These include enhanced training in refusal, red teams of biosafety experts and improved detection systems.
Remember that Anthropic released information about its new Claude Opus 4 and Other surveys of advanced AI modelshave shown that they can take actions that they believe are moral and ethical, but that could compromise the user. For example, emailing government agencies and journalists about suspected wrongdoing by the user.
This model believes it is acting as a “whistleblower”but it may compromise the privacy, security and proprietary information of users and alert authorities where there is no wrongdoing or doubt.
Strong performance in real-world tasks.
The ChatGPT agent has not only performed better in theory, but it has also delivered excellent results on a number benchmarks that simulate real-world knowledge-based work. It achieved a new score of 44.4 on Humanity’s Last Exam using parallel rollout techniques, and achieved 27,4% on the difficult FrontierMath test.
On SpreadsheetBench, it scored 45.5%—more than doubling Copilot in Excel’s performance.
Current limitations and next steps.
Some of the features, such as slideshow generation, may still be in beta. They may have a basic format or differ slightly from in-app previews to exported files. OpenAI is actively working on the next version of this feature in order to improve polish and layout.
With the launch of ChatGPT Agent, users will be able to interact with AI in a new way. They’ll no longer ask questions but instead assign tasks.
OpenAI’s ability to act, reason and produce deliverables will allow users to want AI to not only assist them but also work for them. The company stresses that the agent is still in development, but it sees the launch as a foundation for a future where AI will be more interactive and action-oriented.
VB Daily provides daily insights on business use-casesWant to impress your boss? VB Daily can help. We provide you with the inside scoop about what companies are doing to maximize ROI, from regulatory changes to practical deployments.
Read our privacy policy
Thank you for subscribing. Click here to view more VB Newsletters.
An error occured.
