Agentic search
Agentic search is web search run by an AI agent: the agent plans a query, reads the results, decides what is still missing and searches again until it can complete the task. One task can take dozens of search calls, so speed, cost per call and the quality of titles and snippets matter more to an agent than to a person reading a results page.
The term spread with deep-research and coding agents that call web search as a tool. A log study of 14.44 million search requests that agents sent to the DeepResearchGym API (“Agentic Search in the Wild”, arXiv, 2026) found that over 90% of multi-turn sessions take ten steps or fewer and that, on average, 54% of the new terms in a follow-up query already appear in evidence the agent has read.
Example
Asked which startups a fund backed in 2025, an agent first looks for the fund’s portfolio page, finds three company names it did not know and then runs a separate search for each name to confirm the funding date. The person asked one question, and answering it took four search calls.
How it works in NEEDLE
NEEDLE’s AgenticRare vertical samples rare-word queries from public agentic search logs (DeepResearchGym, OpenResearcher and LRAT), so engines are scored on long-tail queries that agents actually sent. See the live NEEDLE results.
FAQ
How is agentic search different from regular web search?
In regular web search a person types one query and reads a page of results. In agentic search an AI agent writes the queries, reads the titles and snippets itself and decides on the next query, so one task becomes a sequence of searches.
What does agentic search need from a search API?
Agentic search needs low latency and a low cost per call, because an agent may search dozens of times per task. It also needs informative snippets, because the agent decides from them which pages to open.
Related terms
Written by Keenable.
Published
