• ChatGPT Agent handles tasks like bookings and calendars and is more assistant than chatbot.
• OpenAI says it’s still experimental and to use with caution.

OpenAI has rolled out ChatGPT Agent, a tool that the company says can take over digital tasks for users, carrying out actions on a user’s computer. That includes booking appointments, checking the calendar, browsing files, placing online orders, and building presentations, all with minimal user input.

Unlike earlier versions of ChatGPT that mainly focused on writing and answering questions, this works more like a digital assistant. It connects directly to your computer and can access apps, files, and the web to complete tasks. In early demos, the agent helped plan a wedding trip, purchased clothes online, and wrote workplace documents using live data.

OpenAI CEO Sam Altman called it a big step in AI’s ability to act on its own. He said the tool can “think, use tools, take actions, and repeat this process.” But he also warned that the technology is still experimental, stressing that people should be cautious when using it, especially when it comes to sensitive data like emails or bank accounts.

To reduce risk, OpenAI has built in several safety features. The agent needs permission before doing anything permanent – like sending emails or booking travel. It also comes with session logs so users can review what the agent did. There are system-wide limits in place to prevent it from doing anything dangerous, like creating harmful content or accessing restricted parts of the computer. Altman said the safest approach for now is to give the agent access to only what it needs, nothing more.

The company is releasing the tool to ChatGPT Pro, Plus, and Team subscribers first. Enterprise and Education users will get access later this summer. There’s no launch date yet for Europe or Switzerland.

The system is powered by a new AI model that OpenAI trained using reinforcement learning, the same method the company used for its other reasoning tools. The model doesn’t have a specific name, but it was built to use a mix of tools – like a browser, file manager, and terminal – and to handle multi-step tasks. Users can also upload their own data, which the agent can work with directly.

OpenAI combined two previous projects, Operator and Deep Research, to build the new agent. The team behind it includes about 20 to 35 people in product and research. In a demo shown to The Verge, product lead Yash Kumar and research lead Isa Fulford showed the tool planning a night out by checking a Google Calendar and then scanning OpenTable for restaurant options. The user could interrupt the process at any point to adjust the request, like switching cuisines or changing times.

Another demo had the agent create a research report comparing two toy trends. Fulford said she’s used the agent for online shopping and found it more thorough than using previous tools. Kumar mentioned using it for regular tasks like requesting a parking spot each week – something the agent now handles automatically.

Because the agent can access more than just the web – it can use the full computer – it can take on more advanced tasks. That includes putting together PowerPoint slides or analysing Excel files. But with that extra reach comes added risk, with “Watch Mode” purporting to add safeguards. If a user opens a sensitive web page, like a financial account, the agent stops working if the user switches tabs. This is meant to keep users in the loop when there’s potential for misuse.

The agent isn’t especially fast, especially for complex requests. But OpenAI says that speed isn’t the goal. Kumar noted that the team is focused on getting harder jobs right. Fulford added that even if it takes 15 to 30 minutes, it’s still faster than doing the task manually. The idea is that users can start a task, let the agent run in the background, and return when it’s done, presumably so ignoring mis-steps, errors, and hallucinations.

OpenAI has added controls to block the agent from being used in high-risk scenarios, like helping someone make dangerous chemicals or creating biological threats. The company said it hasn’t seen signs of that happening but added the protections anyway. Anthropic, a competing AI firm, introduced similar safeguards earlier this year.

The agent can’t complete financial transactions. Kumar said those functions are off-limits “for now.” Any steps toward that kind of control would require stricter safety checks.

AI agents like this have been a hot topic in the tech world for a while. The idea is to build tools that can go beyond answering questions and actually take action – something like Iron Man’s J.A.R.V.I.S., but more realistic. The term “AI agent” became common in 2023 after Klarna said its own system handled two-thirds of its customer chats in just one month, replacing work normally done by hundreds of people.

Since then, big companies like Amazon, Meta, and Google have all talked publicly about their plans to build similar tools. Google recently hired several leaders from Windsurf, a startup focused on this kind of AI, to help speed up its efforts.

ChatGPT Agent follows OpenAI’s earlier tool, Operator, which could browse the web and interact with buttons and forms. Other companies are moving in the same direction. Anthropic released a similar tool called “Computer Use,” which was built to use a PC like a person would. Several AI companies, including OpenAI, now offer “Deep Research” tools that can pull together detailed reports on just about any topic.

With Agent, OpenAI is trying to show what happens when AI moves beyond writing text and starts managing digital work. But even as the tool gets smarter, the company says it shouldn’t be left to run unsupervised. For now, the message is clear: the agent can be helpful, but it still needs a human nearby.