When I started evaluating browser agents, most of the conversation around me focused on multimodal models, computer-use ...