OpenAI’s New ChatGPT Agent Can Control a Computer and Do Tasks For You

A sleek digital workspace showing a glowing AI hologram manipulating apps and browser windows on a virtual computer screen, cinematic,

When OpenAI released its first conversational chatbot, ChatGPT, in 2022, the world marvelled at its ability to compose emails, write poems, and answer questions. But the company’s latest offering — ChatGPT Agent — marks a more radical departure from what most people think of as a chatbot.

Unveiled this week, OpenAI’s Agent promises not just to generate words but to actually perform multi-step tasks on your behalf, using a virtual computer it controls. In essence, it is no longer simply a conversational assistant. It is an autonomous operator, capable of taking action — planning events, buying ingredients, writing reports, even requesting parking spaces — while you get on with your day.

Beyond Words: Into Action

At a press briefing for The Verge, OpenAI product lead Yash Kumar and research lead Isa Fulford described Agent as a tool that can coordinate an impressive range of complex tasks using a combination of text browser, visual browser, and terminal. The model, unnamed but specifically trained for Agent, uses reinforcement learning to combine the capabilities of OpenAI’s earlier Operator and Deep Research products.

A sleek digital workspace showing a glowing AI hologram manipulating apps and browser windows on a virtual computer screen, cinematic,

OpenAI engineers demonstrated the tool planning a date night: Agent checked the user’s Google Calendar, found a free evening, cross-referenced restaurant options on OpenTable, and even paused to ask whether to include an additional cuisine category before making a reservation.

Another demonstration showed the Agent compiling a research report comparing the popularity trajectories of Labubus and Beanie Babies — gathering and analysing data without a single human keystroke beyond the initial prompt.

A Personal Assistant With a Virtual Computer

Kumar and Fulford emphasised that Agent differs from previous assistants because it has access to a full virtual machine rather than just a browser. That distinction means the Agent can operate with greater flexibility — editing documents, manipulating files, and executing scripts — much as a human would on their own machine.

The result is a tool that feels less like a chatbot and more like a junior colleague working in the background.

One OpenAI employee reportedly uses Agent every Thursday to request office parking in advance, sparing him the Monday scramble to secure a spot — a quaint but revealing example of how even minor tasks can be delegated to the Agent.

Speed, Safeguards, and Limitations

Those hoping for lightning-fast results may be disappointed. In the demo, Agent was not exactly quick. Some tasks reportedly took 15 to 30 minutes to complete — but, as Fulford noted, that is still dramatically faster than a human handling the same tasks end-to-end.

OpenAI seems to have designed Agent for hard tasks, rather than low-latency interactions. Users are expected to set tasks running in the background and return later.

Importantly, the Agent always pauses before executing irreversible actions, such as sending an email or completing a booking. Financial transactions are off-limits “for now,” according to Kumar, and navigating away from certain protected tabs — like banking websites — will automatically halt the Agent’s process.

High Stakes, High Safeguards

The new model powering Agent has heightened capabilities, which raised concerns internally and externally about misuse. OpenAI said it activated its safeguards for “high biological and chemical capabilities,” though it admitted no evidence exists yet that the Agent could actually assist in creating biological or chemical weapons.

Anthropic, a rival company, introduced similar protections when releasing its Claude Opus 4 model earlier this year — underscoring the rising stakes as AI tools grow more powerful.

How To Access ChatGPT Agent

OpenAI began rolling out Agent to Pro, Plus, and Team subscribers this week. Users can activate it by selecting “agent mode” in the tools menu or by typing /agent in ChatGPT. The company plans to make Agent available to Enterprise and Education customers later this summer, though no timeline has been announced yet for Europe and Switzerland.

For now, the Agent remains under close observation, both by OpenAI engineers and by a curious public eager to test its limits.

A Glimpse Into the Future

ChatGPT Agent signals a shift in how we interact with computers: from issuing commands and receiving responses, to delegating tasks and letting AI handle them end-to-end.

In time, tools like Agent could evolve into indispensable virtual employees — managing projects, booking travel, conducting research, and even negotiating deals. But that same power raises hard questions: how much control should we hand over? What responsibilities lie with the human user? And how do we prevent such tools from being exploited maliciously?

For now, OpenAI’s Agent offers a glimpse of what such a future might look like — one where productivity is amplified, but vigilance remains indispensable.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top