Agentic Search

Agentic search equips language models to plan queries, use search and browsing tools, inspect documents, and combine evidence across multiple steps. Common approaches include lexical and dense retrieval, direct interaction with raw corpora, multi-agent exploration, and training with supervised trajectories or reinforcement learning. The field evaluates these systems on multi-hop question answering, deep research, scholarly discovery, and domain-specific tasks, increasingly measuring evidence coverage, citation faithfulness, and cost alongside answer accuracy. A central tension is how to make long searches more reliable and complete without wasting tool calls or context; stronger reasoning does not always compensate for poor retrieval, weak evidence tracking, or ineffective stopping.