Lean Search

Overview

Web research has two distinct costs, and they need different fixes:

• Main context-window bloat — raw search results and fetched pages pile up in the conversation, degrading quality and pushing toward limits. Fix: isolation (do the searching inside a subagent that returns only a conclusion).
• Total billed tokens — every token in every context bills. Fix: frugality (search before fetching, fetch narrowly, never fetch twice).

Isolation and frugality are different dials. A subagent protects the main context but does NOT lower the bill — both contexts bill, and an unnecessary subagent adds a whole extra context. So: isolate when the research is big, stay frugal always.

NOT for the heavyweight deep-research skill's job (multi-source, adversarially verified reports) — this is the everyday "I need a fact / a doc / a few sources" path.

How the tools actually bill

The flow

1. Search first — read the snippets

Run WebSearch before any fetch. Scope it with allowed_domains when you know the authoritative source (e.g. docs.python.org, a specific vendor's docs) so you get fewer, better results. Read the snippets. If they answer the question — stop. This is the "only look at headers" path, and it's the cheapest one. Most lookups end here.

2. Need a page? Pick inline vs. isolated — by expected fetch count

• 1 known URL, 1 specific question → fetch INLINE. Spawning a subagent for a single page adds a whole context to the bill for no context-window benefit worth it. Just WebFetch with a tight prompt (step 3).
• Multiple pages / exploratory / "research X" / unknown how many sources → ISOLATE. Dispatch ONE subagent to do all the searching and fetching; it returns a short, cited synthesis. The main thread pays only for that summary (step 4).

Rough line: ≤1–2 fetches → inline; more, or unknown → subagent.

3. WebFetch discipline (always)

• One specific extraction prompt per fetch. Ask for the thing, not the page: "Return only the default value of timeout and the version it changed in" — not "summarize this page." The prompt is what bills into context.
• Don't fetch what a snippet already gave you.
• Don't fetch the same URL twice (it's cached ~15 min, and you don't need it again).
• Never ask it to dump the whole page or "read the full doc into context."

4. Subagent discipline (when isolating)

Isolation alone protects the context window but can quietly run up the bill if the subagent fetches everything. Instruct it to be frugal too:

• Snippets-first; fetch a page only when snippets fall short.
• Cap the fetches (e.g. "fetch at most 3 pages").
• Tight extraction prompt per fetch.
• Return a short, cited answer (e.g. "≤150 words + source URLs"), not a transcript of what it read.

5. Reuse, don't re-run

Already have search results or a fetched extract? Use them. Re-searching to "double-check" re-bills for information you already hold.

Red Flags — rationalizations to reject