We believed that Chatgpt was just a very capable chatbot. Openai has just turned it into something very different: a real agent

We have been talking about artificial intelligence agents for a long time, but Openai has just converted that conversation into something much more tangible. The company has presented Chatgpt Agent, a function that turns its popular assistant into something more autonomous: now it is able to execute complex tasks using a virtual computer, with tools that allow you to navigate, program or even make decisions.

From Agent Operator. At the beginning of the year it presented Operator, a tool that allowed ChatgPT to interact with web pages. Then Deep Research arrived, focused on writing long reports from multiple sources. The background idea was clear: go beyond the conversation and approach real tasks. What has been presented today is something like a tool that unifies all these previous advances.

During the demonstration, those responsible for the project raised a daily situation: organizing a trip to attend a wedding. The agent was able to understand the context, find hotels, propose gifts, take into account the weather, the clothing code and even remember that a suit had to be bought. He did it by analyzing the message, accessing the web and acting step by step, as a person would. The difference is that everything happened within Chatgpt, without the need to alternate tabs or give instructions one to one.

A virtual computer for AI. The key is that the agent is not limited to responding to text: it operates within a kind of virtual computer that Openai has given access. You can use a text browser to read pages quickly, a visual browser to interact with buttons and forms, and even a terminal to run commands, generate code and manipulate files. You can also work with spreadsheets, presentations, and access services such as Google Drive, Calendar or Github if the user authorizes it.

What is under the hood? The model that drives chatgpt agent (specifically developed for this function, although without official name) was trained with complex tasks that required to combine multiple tools. Openai used reinforcement learning, the same approach that you already use in its reasoning models, to teach you to choose when to use the browser, the terminal or an API. The idea was to develop a solution capable of accurately deciding how to act based on each objective.

In development.

Images | OpenAI

In Xataka | Goal is in a hurry to lead the AI that has done something unusual: it is building a data center in tents