The bitter lesson of browser agents

# Summary Anthropic's Browser Use tool evolved from restricting AI models to predefined actions and page states to giving them direct access to the Chrome DevTools Protocol (CDP), allowing models to write their own code for both actions and observations. This shift reduced token usage by 60-66% while maintaining perfect task completion rates, but revealed that the real limitation wasn't the action space—it was the developers' assumptions about what information the model needed to see. The key insight is that removing human-imposed constraints on both what models can do and what they can perceive leads to more efficient and reliable automation.