Agentic Search explained: a 4-search loop costs about 5x one search — and queries leak to more companies

adnan_hashmi · x · 2026-11-12

Paolo Perrone explains agentic search: a model searching in a loop — writing a query, reading results, and picking its next move (search again, open a page, or answer). Coding agents use it to find code behind bugs; research agents use it to gather sources.

Key points:

The loop's value is question decomposition: in the classic Self-Ask example ("US president when superconductivity was discovered"), GPT-3 got it wrong with no search (FDR) and one search didn't help; splitting into "when was it discovered (1911) → who was president in 1911 (Taft)" answers correctly
Cost comes mostly from resending context: every call carries the question plus all prior results, so a 4-search loop costs about 5x one search
Privacy: where the model-written query lands depends on the search API — a self-indexed API reaches 1 company (e.g. Brave keeps API queries up to 90 days), a scraper-based API reaches 2 companies, and a model provider's built-in search reaches the index provider it buys from

Takeaway: the loop often pays off when the first search misses, but budget for multi-turn context resending.

Original post →

Agentic Search explained: a 4-search loop costs about 5x one search — and queries leak to more companies

More from coding & agent