You give Fara a goal in plain English, such as "book a table for two at a sushi place on Friday at 7pm", and it takes over the browser. It looks at screenshots of the page, the same view you have, decides the next move, then clicks, types, scrolls, and searches until the task is done. It does not read the page's hidden code; it navigates by sight, which means it copes with any site a human can use.
The design choice that matters for trust: it is trained to stop at what Microsoft calls critical points. Before entering personal details, typing card numbers, signing in, or hitting any submit button it cannot undo, it pauses and asks you instead of guessing. On public web-agent tests it completes 89.3 percent of WebVoyager tasks and 72.3 percent of Online-Mind2Web, the strongest open-weight results in its class.