How do AI agents browse the web?

AI agents browse the web by calling a tool — a function exposed to the model that fetches a URL, cleans the response, and returns it as text the model can reason over. In 2026, the dominant pattern is MCP (Model Context Protocol) servers like AgentFetch, which expose fetch_url, extract_json, and search tools to Claude Desktop, Cursor, Cline, and LangChain agents through a standardized JSON-RPC interface. The agent decides when to call the tool; the MCP server handles the dirty work: following redirects, retrying on 429/503, stripping nav/footer/ads, converting HTML to markdown, and handling encoding quirks. Raw fetch() or requests.get() is rarely sufficient — modern sites return JavaScript shells, Cloudflare interstitials, or 30KB of boilerplate around 500 words of content, all of which waste model context. Production agent stacks therefore lean on a fetch layer (AgentFetch, Firecrawl, Jina Reader) that returns ~80% smaller, semantically clean output. Browser-automation tools (Playwright, Browserless) are reserved for pages that require execution, login, or interaction.