Browser agents are the use case where RL makes the most sense - the reward signal is obvious (did the task get done or not) and the action space is bounded. Curious how you handle the credit assignment problem across multi-step navigation though.