Arxiv paper proposing a practical architecture for building browser-based AI agents. Covers system design patterns, sandboxing and security considerations, and how to make agents robust to real-world web complexity. Includes evaluations on common browsing tasks and discusses deployment tradeoffs and open problems.