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

web_search(query, num_results?, topic?, search_depth?)

Arg	Values	Notes
query	string	Batch many in one call with |||: "a ||| b ||| c". Each runs in parallel.
num_results	1–20, default 5	Per query, not a batch-wide cap.
topic	general (default), news, finance	news for breaking/current events, finance for markets, tickers, filings.
search_depth	basic (default), advanced	advanced is slower and more thorough — reserve it for hard or ambiguous questions.

Each query returns JSON: a synthesized answer, then results[] of {title, url, snippet, score, published_date}, plus any images[] of {url, description}. Batched calls return {batch_mode, total_queries, results[]} — one block per query, in order. A failed query comes back as {query, success: false, error} without sinking the rest.

The snippet is already query-focused — treat it as the cheap answer and escalate to a fetch only when it's thin or contested. score ranks relevance; published_date is your recency signal.

scrape_webpage

scrape_webpage(urls, include_html?)

Arg	Values	Notes
urls	string	One URL, or several comma-separated: "https://a.com, https://b.com". Batch in one call.
include_html	bool, default false	Leave off — markdown is what you want almost always.

Returns clean markdown as content (with title, content_length, metadata). Multi-URL calls return {total, successful, failed, results[]} with per-URL success, so one bad page doesn't sink the others. Timeouts are retried internally. For GitHub repos or files, skip this and use gh via bash instead.

Querying

• Keep each query to a handful of strong keywords. Vary the angle across the batch, not just the wording: claim plus counter-claim, official source plus independent coverage, and a dated query for anything time-sensitive.
• Set topic per query: news for "what happened", finance for numbers and markets, general for the rest.
• Reach for search_depth="advanced" only after basic comes back thin — it costs latency.
• For scholarly sources, point queries at the literature (… systematic review, site:arxiv.org, journal names) or load the openalex-paper-search skill for a real paper index.

Example fan-out:

web_search(
  "EU AI Act high-risk obligations enforcement date ||| EU AI Act 2026 compliance deadline ||| AI Act general-purpose model rules industry criticism",
  search_depth="advanced",
)

Patterns that scale

• Fan out, don't loop. One batched call beats N sequential ones — both tools parallelize internally.
• Dedup before you fetch. The same URL surfaces across query variants. Collapse on url, keep the richest snippet, then spend a scrape.
• Snippet-first. Fetch only the URLs whose snippet didn't close the question.
• Disk is your memory. Beyond a quick lookup, write scraped pages and extracted notes to a working dir and read them back selectively, rather than holding raw page text in context. This also makes long jobs resumable.
• Big pools → batch and delegate. Reading dozens of pages: scrape in batches, write each to disk, process in chunks — or hand subtopics to parallel subagents — instead of one giant context. Only the top-level agent can hand off to subagents; if you are already a subagent, chunk through the pool yourself via disk.

Full mechanics in reference/scaling.md. End-to-end recipes in reference/recipes.md.

Output

Every factual claim traces to a page you read or a snippet you saw.

• End with a Sources section of markdown links: [Title](url).
• Cite inline where it matters; quote directly only when exact wording carries weight.
• Flag disagreement between sources instead of smoothing it over. Say so when the evidence is thin or absent — "no solid source found" is a valid answer.
• Never invent a URL. If you didn't open it, don't cite it.

Pitfalls

• A high score means relevant, not correct — read before trusting.
• Snippets go stale. Check published_date on time-sensitive facts and prefer topic="news".
• Paywalled pages scrape thin — fall back to the snippet or a secondary report, and note the gap.
• Read the error field on a failed scrape instead of retrying blindly; it tells you why.
• GitHub, large PDFs, and login-walled pages are poor scrape targets — use gh, a PDF route, or a different source.