New Agent Mode lets ChatGPT act on your behalf

OpenAI has introduced a major new capability to its flagship product: ChatGPT Agent Mode. Officially rolled out on 17 July 2025, the new feature enables ChatGPT to complete complex tasks from start to finish using its own virtual computer.

Now available to Pro, Plus, and Team users, Agent Mode gives ChatGPT the power to do more than just chat — it can take action on your behalf. That means it can browse the web, run code, access APIs, and even log into your Gmail (with your permission, of course) to help complete tasks. Whether it’s pulling together a spreadsheet or building a slideshow, Agent Mode handles the heavy lifting. You can turn it on anytime through the tools dropdown in any ChatGPT conversation.

Agent Mode combines the strengths of OpenAI’s earlier tools — Operator and deep research — into a unified agentic system that balances reasoning and real-time action. OpenAI noted that the agent always asks for permission before performing any consequential action and allowing users to pause, interrupt, or take over at any time.

The system operates using its own virtual computer, with access to a range of tools such as a graphical web browser, a text-based browser, a terminal, and API connectors. For added flexibility, users can take over the browser to log in securely, enabling the agent to access gated content or carry out tasks on authenticated platforms.

OpenAI says Agent Mode significantly outperforms previous models in benchmarks covering data science, investment banking tasks, and spreadsheet editing. It also includes safety layers to prevent misuse, such as protections against prompt injection and explicit user confirmation before taking sensitive actions like purchases.