The agent successfully assembled a comprehensive explanation of BlockRun by combining web search results, prior knowledge, and attempted page fetches—most of which returned incomplete HTML. The site's pages did not render fully in the agent's text-based fetch, forcing reliance on search snippets, GitHub documentation, and prior knowledge to construct the final answer. The agent filled gaps intelligently but had to work around the site's poor machine-readability.

• ›Step [3] (pricing search) and step [10] (site search) returned snippets and URLs that directed the agent to blockrun.ai/pricing and confirmed key differentiators (x402 protocol, USDC payments, routing), but the actual fetches of those pages ([4], [7]) returned truncated/unrendered HTML with no visible text content.
• ›The agent compensated by citing GitHub repositories (awesome-blockrun, blockrun-llm, blockrun-mcp) discovered through search [3] and [10], leaning heavily on prior knowledge (40% sourced from prior_knowledge per metadata) to fill in details about pricing models, agent autonomy, and the x402 protocol.
• ›The site is not agent-ready: its Next.js-rendered pages returned only boilerplate HTML and script tags with no semantic markup, structured data, or fallback text. The agent had to rely on web search to discover and validate the site's own content rather than fetching it directly.
• ›Friction was 'light_bridge'—the agent succeeded despite the navigation barrier by using search and prior knowledge as a bridge, not by the site providing clear, machine-readable surfaces.