New ChatGPT Agent: A Leap Towards Autonomous Task Execution for Users

The startup OpenAI has introduced a new universal AI agent within ChatGPT, designed to carry out a variety of computer tasks on behalf of users.

The company claims that this agent can automatically manage the user’s calendar, generate editable presentations and slides, and execute code.

The ChatGPT agent merges several functionalities from earlier agent solutions. It includes features such as the Operator’s ability to click through websites and the Deep Research function, which collects information from numerous sites and offers a concise analytical report.

Users can interact with this tool using natural language during a conversation with the chatbot.

Initially, the AI agent is available to Pro, Plus, and Team subscribers. To activate it, users must select «agent mode» from the ChatGPT tools dropdown menu.

OpenAI asserts that the new ChatGPT agent significantly outperforms other available solutions. It can utilize ChatGPT connectors to link applications like Gmail and GitHub for retrieving essential information and responding to queries. Additionally, it has access to a terminal and can employ APIs.

The digital assistant’s capabilities cover tasks such as planning and purchasing ingredients for a Japanese breakfast for four and analyzing three competitors followed by creating a presentation.

OpenAI highlighted that the underlying model of the tool exhibits superior performance across various benchmarks. In Humanity’s Last Exam—an extremely difficult test featuring thousands of questions spanning over a hundred subjects—the ChatGPT agent achieved a score of 41.6%, approximately double the scores of o3 and o4-mini.

In one of the most challenging mathematical assessments, FrontierMath, the neural network scored 27.4%, surpassing the previous record set by o4-mini which was 6.3%.

The startup emphasized that security considerations were paramount in the development of the ChatGPT agent, as the enhanced capabilities could pose risks if misused by malicious actors.

According to a report, the model has been classified as having «high capability» in the domains of biological and chemical weaponry. This indicates that it could potentially enhance existing methods of inflicting serious harm. OpenAI clarifies that while there is no direct evidence of such a threat, it is adopting a preventive approach and implementing additional safeguards. These measures include:

It should be noted that in July, OpenAI revised its security protocols to protect intellectual property from corporate espionage amid concerns regarding theft by Chinese competitors.

Previously, ChatGPT was trained to connect to a greater number of internal sources, allowing it to access contextual information in real time.