Browser Agents

Browser agents let AI systems use websites as tools. They navigate live pages, click and type, search across sources, collect evidence, and complete web workflows that would otherwise require a person at a browser.

The problem is larger than opening pages. A useful browser agent has to understand changing interfaces, plan over multiple steps, recover when a path breaks, verify what it found, and know when it has enough evidence to stop.

Our work studies the interface between general models and real browsers: what the model should decide, what the browser layer should expose, and how much structure helps before it starts hiding useful state.