Browser agent

Browser agents let a model use websites that have no API. The agent sees the page, either as a screenshot, as structured text extracted from the page, or both, then chooses actions like clicking a link or typing into a field.

They are useful for research, data collection, testing your own web apps, and chores like filling in repetitive forms. They are also slower and less reliable than calling an API directly, because websites change and pages can be ambiguous.

Browsing is one of the riskiest things an agent can do. Every page it reads is untrusted content that may contain prompt injection, and a logged-in browser can take real actions with your accounts. Limit what it can access and require approval for anything consequential.