What is browser agent?
A browser agent takes a goal in natural language, observes the current page, decides the next action, performs it, and repeats until an external success condition holds or a step budget runs out. It differs from a script in that it re-decides at every step, and from a chatbot in that it changes state on a real site.
Why it matters
Adapting per step is what lets an agent handle a page it has never seen — and it is also what makes it expensive, because every step is a model call. Compiling a successful trace is how a browser agent stops paying for the same reasoning twice.