ChatGPT incorporates Agent with autonomous capabilities to execute complex tasks

OpenAI announced the launch of ChatGPT Agent, an agent integrated into ChatGPT that allows the assistant to execute complex tasks from start to finish using its own virtual computer, but always under user supervision. This new capability can perform actions such as analyzing calendars and preparing reports on upcoming meetings, planning and purchasing cooking ingredients, or analyzing competitors and creating presentations.

The system combines three previous capabilities: Operator's ability to interact with websites, Deep Research's analysis capabilities, and ChatGPT's conversational intelligence. The tool navigates web pages, filters results, executes code, performs analysis, and generates editable documents such as presentations and spreadsheets.

ChatGPT Agent operates through a virtual browser that can alternate between reasoning and action to handle complex workflows. The tool includes a visual browser that interacts with graphical interfaces, a text-based browser for simpler queries, terminal access, and direct API connection. It can also integrate with applications like Gmail and Github through connectors.

Users maintain full control over the process. ChatGPT Agent requests permissions before performing important actions and allows interrupting, taking control of the browser, or stopping tasks at any time. For critical actions such as sending emails, it requires active user supervision.

The functionality is available to Pro, Plus, and Team plan users by selecting "agent mode" from the tools dropdown menu in the ChatGPT interface. ChatGPT Agent is not available in Europe or Switzerland due to regulatory reasons.

OpenAI reports significant improvements in benchmark evaluations. On Humanity's Last Exam it reached 41.6 points, and on FrontierMath it achieved 27.4% accuracy. In data science tasks it surpassed human performance by a considerable margin, and on SpreadsheetBench it achieved 45.5% compared to Copilot in Excel's 20%.

ChatGPT Agent includes security measures against adversarial manipulation and prompt injection attacks, where malicious instructions hidden in web pages could attempt to modify its behavior. OpenAI implemented explicit user confirmation for important actions, active supervision for critical tasks, and proactive rejection of high-risk operations such as bank transfers.

The company has implemented the strictest security measures developed to date for ChatGPT Agent, applying special controls due to the system's advanced capabilities.