Search Results for "scrolling"

...Each browser state is converted into a numbered table of currently available interactive elements. A policy chooses both an operation and a compatible target in a single network round trip. Supported actions include clicking, typing, selecting, scrolling, waiting, finishing, and reporting that a task is blocked. A small language model is invoked only when text must actually be generated for a typing action. The agent connects to Chrome through Browser Harness and provides an inspector showing elements, probabilities, decisions, and executed actions. Its design avoids site-specific scripts and instead relies on fresh observations of the live page.