Kushan Raj argues that a browser agent can struggle even with a capable model if its environment offers fragmented page information and weak feedback about actions.
He demonstrates a prototype that combines a compact representation of the full page with screenshots and tracks changes, obstacles and failed interactions. He contrasts its behavior on a few tasks with another agent.
Raj describes open-sourcing and an API as possible future directions. The examples illustrate his hypothesis but do not establish a general controlled performance comparison.
Watch on YouTube
