OpenAI, one of the key AI companies, has made a big step towards autonomous agents that can act almost like virtual assistants. It introduced ChatGPT Agent — tool, can perform work on behalf of the user using its own «virtual computer».
ChatGPT Agent is based on a new model created specifically for this product. Based on the user’s request, it can check websites, filter results, prompt the user to log in when necessary, run code, perform analysis, create spreadsheets and PowerPoint presentations, and more. It can perform a variety of tasks, such as viewing your calendar and preparing a summary of upcoming client meetings, making a breakfast plan for the whole family with the right groceries, or creating a presentation based on competitor analysis.
Since ChatGPT agents do all their work using their own virtual computer, they will have all the context they need to complete the task. For example, an agent can visit a website using a browser, download a file from the site, process the same file by running a command in the terminal, and then view the result in a visual browser.
The model behind ChatGPT Agent does not yet have a specific name. It was trained on complex tasks requiring multiple tools, such as a text browser, a visual browser, and a terminal where users can import their own data. It was trained using a reinforcement learning methodology — the same one that OpenAI uses to train logical reasoning models. ChatGPT Agent combines the capabilities of the two previous solutions OpenAI — Operator and Deep Research. And it was the teams that developed these tools that merged to create the new agent. The new combined team consists of 20-35 people who are engaged in both research and product development.
During the demonstration, they showed several potential scenarios for using ChatGPT Agent. For example, to plan a date, it can independently check the available time in Google Calendar, find a restaurant via OpenTable, and book a place for dinner. The user can intervene in the process, for example, by adding new restaurant categories. ChatGPT Agent is also capable of preparing a full report on the popularity of Labubus toys compared to Beanie Babies.
The new tool can also be used for online shopping — it works more accurately and thoroughly than Operator alone. In addition, the agent can be used for small routine tasks, such as applying for a parking space at the office every Thursday.
ChatGPT Agent has access not only to a browser but also to a full-fledged virtual computer, which allows you to significantly expand the set of functions. However, the tool is not very fast. If the task takes 15-30 minutes, it is still a significant time savings compared to manual execution. You can «run the agent in the background» and return to the results later.
To prevent unwanted actions, the agent always asks for permission before sending an email or making a reservation. And if the user visits financial websites, the observation mode is activated: the agent works only in the tab in which it was launched and stops if the user switches to another.
OpenAI claims that ChatGPT Agent demonstrates excellent results in various tests that evaluate the ability to browse the web and perform real-world tasks. For example, it achieves 27.4% accuracy in FrontierMath, significantly outperforms human performance in DSBench data processing tasks, and scores 65.4% in WebArena.
OpenAI starts gradual deployment of ChatGPT Agent for Pro, Plus, and Team users. It can be found in the tools menu or called with the /agent command. Later this summer, the agent will also be available for corporate and educational ChatGPT subscriptions. However, there is no clear launch date for the countries of the European Economic Area and Switzerland.
Spelling error report
The following text will be sent to our editors: