OPENAI Lance Chatgpt Agent, a virtual computer that does everything for you

After the Chatgpt revolution in 2022, the Chatgpt Agent Revolution in 2025?

Thursday, July 17, Openai unveiled a major novelty in the generative artificial intelligence sector: he will integrate a virtual mini-PC into his chatbot capable of doing things for you. Its concept is quite simple: tell Chatgpt to go book your vacation after checking your calendar and controlled your budget, it will do it by logging into comparison sites, your bank account and creating tracking documents. The tool can solicit you from time to time to request confirmation or password but is designed to be autonomous.

Chatgpt Agent replaces Operator, the experimental web browser integrated in Chatgpt since the beginning of 2025. The company explains to Numerama to have merged its Operator and Deep Research teams (its model that does long research) to design this new tool that has revolutionized modern computer. Chatgpt Agent is available today for pro subscribers and more, but will arrive with delay in the European Union.

Chatgpt Agent: The start of the end for web browsers?

Chatgpt Agent is presented as a “universal agent” capable of performing precious tasks for the user.

Permanently accessible from the main view in the manner of Deep Research, or trigger with the ” /agent” command, it generates a window in which things happen. We see Chatgpt moving the mouse cursor and launching applications, with text that indicates on the screen what it is doing.

Compared to the Deep Research mode, which was content with long searches synthesized in a single document, Chatgpt Agent has many tools. He can explore the web of course, but also create files, connect to accounts, control a terminal, generate slideshow, etc.

One of its particularities is not to depend on the extensions already available in Chatgpt (Google and Microsoft for example) and to be able to connect to any site. When he needs an identifier and a password, he leaves the hand to his user. OPENAI explains to Numerama having designed Chatgpt Agent so that he does not do everything alone and asks his user what he wants. Another peculiarity: a cursor allows you to go back in time to see everything that Chatgpt has done. We can thus understand all his reasoning.

Among the examples shown in Numerama by Openai: shopping, filling complicated forms (hello taxes), making reservations, checking your calendar and finding availability for restaurants, organizing a schedule, completing a document with data from the web, etc. It is obviously possible to ask Chatgpt Agent to do several things at the same time: it then acts as Super Deep Research.

Let an AI Check your personal data: a new line crossed?

Obviously, the security question arises. With Chatgpt Agent, Openai unveils his vision of the future: computers will authorize themselves. One of the next steps could be the launch of a browser with the integrated tool, before going even further with a new machine without equivalent.

To reassure his intentions, Openai makes several promises:

• The interaction is designed to be collaborative, as if you worked with a person. The agent requests you whenever he has to do something important and can be interrupted at any time. It is not designed to replace you.
• When you have to enter a password, Openai lets the user check the virtual computer. Chatgpt does not record what you write and does not make a screenshot.
• The preview window provides total transparency to the user. Unlike Deep Research, Chatgpt Agent is not satisfied with a summary of what he has done: he shows it in real time.
• Before taking critical actions (such as submitting forms or making a purchase), the agent will always request confirmation from the user.
• The agent will refuse to carry out a large category of tasks defined by OpenAi policy, including harmful actions.
• Chatgpt Agent has no memory. Unlike the rest of the chatgpt, he does not remember what he did in previous sessions (at least at launch).
• Finally, Openai has implemented measures to detect the injection of malicious prompt, note the reliability of a site or not to navigate itself on too sensitive sites (messaging, banks, etc.). He indicates that the hallucination rate is extremely low with this mode.

When will Chatgpt Agent arrive in the European Union? Our interlocutors at Openai explain that its offbeat launch is only provisional, the time to check the legality of the system. Quite quickly, the tool should appear in France. It is already available in the United States, Switzerland or the United Kingdom (and with a VPN). Next step for the company: reveal the long-awaited GPT-5.