Research & Recon Agent

Two things make web research untrustworthy when a model does it. First, citation theatre: quoting a source from its search-result snippet without ever opening it, so the citation looks rigorous and isn’t. Second, brittleness: half the useful web now sits behind bot-walls and JavaScript challenges, and an agent that treats a blocked page as “read” will confidently cite an empty screen. The brief had to be genuinely grounded and had to survive a hostile web.

The agent plans sub-questions, searches via a keyless endpoint, and reads pages through a fetcher that extracts readable text — and, critically, raises an error when a page is a bot-wall or empty. That turns a dead end into a signal the agent adapts to: it picks another source rather than citing a blank. Only pages it successfully read can be cited, and it is pushed to consult several sources before it writes, noting disagreement rather than smoothing it over.

A runnable agent. Asked to explain how agentic RAG differs from classic RAG, it searched, hit Cloudflare challenges on two candidate sources, recognised them as unreadable, re-routed to reputable engineering sources it could actually parse, and wrote a tight three-paragraph brief citing only those. The recorded run below shows the whole thing — including the recovery from the blocked pages.