The agent assembled a comprehensive explanation of Royal Holloway University by combining direct site fetches with web search fallbacks. The site's navigation was partially broken (404s on /study and /undergraduate) and content was truncated during fetches, forcing reliance on prior knowledge and external search results rather than direct site content. Despite these friction points, the agent successfully synthesized enough information to explain the institution's mission, audience, pricing, and differentiation.

• ›Steps [1, 3, 5, 7, 10, 14] returned structured HTML but were truncated in the agent's view; the agent could confirm pages existed (200 status) but had to reconstruct substance from metadata, page titles, and external sources rather than body content.
• ›The site's URL structure is inconsistent and brittle: /study, /undergraduate, and /study-with-us all returned 404s, but /courses, /about-us, and /fees-and-funding worked. The agent had to guess at correct paths rather than discover them through navigation.
• ›Direct site content (fees pages, about us, course listings) was present but inaccessible to the agent in usable form due to truncation. The agent filled gaps with web search results and prior knowledge, indicating the site publishes information but not in a machine-readable, traversable structure.
• ›Pricing information existed on the site but was fragmented: undergraduate fees on one page (/studying-here/fees-and-funding/undergraduate/tuition-fees/), postgraduate on another, with no unified pricing view. The agent had to piece these together.
• ›The About Us page loaded but contained no visible content (marked [truncated, 4400 more chars]) twice, indicating either heavy JavaScript rendering or poor content structure for automated access.
• ›The agent successfully cited 8 of its 14 tool calls as sources in the final response, but 6 of those were direct site fetches that mostly confirmed URLs and metadata rather than providing readable body content; the heavy lifting came from web search snippets and prior knowledge (56% prior knowledge noted in run metadata).