Agentic search is an ambiguous term that covers three distinct implementation patterns. Retrieval-centric approaches build high-quality search so agents can find relevant context on their own, but are limited by imperfect retrieval quality. Harness-centric approaches put the agent in charge of exploration, using relevance feedback and external judges to steer it toward correct results — effective but token-expensive. Model-centric approaches fine-tune an LLM specifically on search tool-calling traces, enabling efficient, domain-aware retrieval without repeating costly exploration steps. Each pattern has trade-offs, and the field is converging toward hybrid designs borrowing patterns from coding agents.
Table of contents
Retrieval-Centric ImplementationHarness-centric implementationsModel-centric approachThe alchemy of ambiguity544 Impressions