How much of your agent's context gets eaten just driving a browser?

Something I keep hitting with AI agents that browse: the browsing itself eats the budget. In an MCP client like Claude Desktop, every look at a page sends back the whole tree, so after two or three pages the context is full and the agent slows to a crawl. A simple task turns into a dozen round trips.

I want to hear how others handle this:

- If your agent drives a browser, how many pages before the context gets tight?

- Do you trim the page before feeding it back, cap it, or just eat the tokens?

- For Claude Desktop specifically (no shell, so no CLI tools), what are you actually using today?

Full disclosure, I built Reflex to solve this for myself, so I am not neutral. But I am mostly curious where the token and round trip cost hurts you most, and what you have already tried.