Agentic Search

This chapter provides a comprehensive survey of agentic search -- the paradigm in which AI systems actively pursue answers through multi-step information-seeking strategies rather than passively returning ranked documents. We cover the full spectrum from foundational retrieval-augmented generation (RAG with dense retrieval, REALM, RETRO, Self-RAG, GraphRAG) through tool-augmented retrieval (WebGPT, Toolformer, ReAct), multi-hop reasoning search (IRCoT, FLARE, DSPy), agentic web browsing (WebArena, Mind2Web, GAIA), search with planning and tree search (Tree-of-Thought, MCTS-LLM, AlphaProof), search in code and mathematics (AlphaCode, SWE-Agent, formal theorem proving), to fully autonomous deep research systems (STORM, OpenAI Deep Research, Perplexity). Throughout, we emphasize the underlying computational principles -- search as sequential decision-making, the explore-exploit tradeoff, verification-guided search, and test-time compute scaling -- that unify these diverse approaches.