In September 2025, I tried ChatGPT Agent to see how well it could perform a task inside a website rather than merely describe the steps. The feature’s July 2025 launch announcement covered browser interaction and research tools. Plan availability and product names can change after that launch.
My experiment was to have it browse videos on my Bilibili account and organize a few favorites around my interests. It took longer than doing a small task myself, but seeing it move through the pages and work toward the result was interesting.
Logging in involved handing browser control back to me. Completing a task also depends on the website, available tools, and permission to take the requested actions. Being able to use a browser does not mean an agent can successfully perform every operation on every service.
I can imagine useful applications for collecting information, comparing pages, and organizing content. If an action publishes something or changes another person’s experience, review the result and authorize it deliberately. For schoolwork, assistance with research or explanation should follow the course’s rules; disguising automated work as your own learning defeats the point.
What appealed to me was the move from answering questions to working with an interface. It was an imperfect but memorable first look at that capability.
Adapted from the original Chinese article, published on September 30, 2025.
Comments NOTHING