Search

Two tools cover the whole loop: web_search finds pages, scrape_webpage reads them. Most questions resolve from search snippets alone — only fetch a page when a snippet can't settle the point.

The loop

1. Frame — restate the question as 2-4 concrete things you need to know.
2. Search — fan out one batched web_search, one query per angle (see Querying).
3. Triage — scan titles, snippets, and scores; pick the few URLs worth reading; drop duplicates.
4. Fetch — scrape_webpage the survivors in a single batched call.
5. Extract — pull the facts you need out of the returned markdown. You are the extractor; there is no separate extraction model to call.
6. Answer + cite — synthesize, then list every URL you actually used as markdown links.

Stop the moment the question is answered. Don't fetch a page whose snippet already gave you the fact.

Tools

web_search