Yeah, OpenClaw agents have a full set of tools to interact with a browser in arbitrary ways. My idea was to instead give it a tool for a browser wrapper with a limited API surface. And that tool could use LLMs internally in specific contexts.