The quiet failure
A site is offline or hard-blocked. Ordinary tools hand back a challenge page as if it were content; Search returns an explicit failure naming both attempted routes — shown here with a real captured error, correlation id included.
The research question
"Summarise what this page says." — but tonight the source is offline or hard-blocked. What the agent does next decides whether anyone can trust the answer.
The failure nobody sees, next to the failure on record
The dangerous case isn't an error — it's a 200 with the wrong body. The site serves its anti-bot challenge page, an ordinary tool delivers it as "content", and the agent summarises it. Search detects challenge and error content and never serves it as a page — the failure comes back explicit, naming both attempted routes. Left pane is an illustrative scenario; right pane is a real error captured 8 Aug 2026 from Claude Code over MCP (a retired host whose TLS certificate no longer validates).
// generic fetch — the site answers with its challenge page fetch("https://forum.example/thread/4471") <!DOCTYPE html><html><head> <title>Just a moment...</title> "Checking your browser before accessing forum.example. Please enable JavaScript and cookies to continue." ✗ 200 OK — the agent "summarises" a checking-your-browser page as the source. No one can tell.
// captured 8 Aug 2026 — error content detected, never served as a page fetch("https://deno-was-retired.example.dev/changelog") FAILED: Error: fetch failed for https://deno-was-retired.example.dev/changelog: direct (blocked/error content: {"error": 500, "correlation_id":"15a074e1ae81"}); stealth (stealth 502: Page.goto: net::ERR_CERT_AUTHORITY_INVALID …)
The answer the agent can now give
Nothing was summarised from nothing — the honest answer is short, specific, and actionable.
I couldn't read the source. Both routes failed — direct fetch hit error content (HTTP 500), and the full browser hit an invalid TLS certificate. No summary was produced. The failure is recorded (correlation id 15a074e1ae81); the domain is on the investigation queue. Want me to try an archived copy, or another source?
Measured improvement, and where this came from
Without Search: a confident summary of a checking-your-browser page, presented as the source — a wrong answer no one can detect. With Search: zero fabricated answers — an explicit failure with both routes named, on the record, with a next step.
client Claude Code over MCP 10:03:53 fetch_results → FAILED, both routes named, verbatim — from the queued-reads run trace correlation id 15a074e1ae81 ties the attempt to service logs recorded workspace activity + stealth_failures queue with failure count