Three Kinds of Agentic Search — Retrieval, Harness, or Model

Doug Turnbull argues that "agentic search" is a confusing term that actually refers to three distinct implementation patterns. All three address the same core problem: agents don't know how to find the right answer in your domain — they make false assumptions about what your users consider relevant, and they need domain context to steer them correctly.

Retrieval-centric puts search in charge. Build high-quality search that solves any conceivable query from an agent, letting search "lead the agent by the nose" toward relevance. The big downside: search is hard, almost nobody builds Google-quality results, and agents trust search — so even simple distractors in retrieval can confuse reasoning. Classic RAG also struggles when answers don't look like answers (e.g., a book synopsis that opens with plot, not the title).

Harness-centric puts the agent in charge. Give it stripped-down retrieval primitives (BM25, filesystem tools) and let it explore. Inject domain knowledge via a judge that labels results as relevant or not — "relevance feedback on steroids." On the ESCI dataset, agentic search with oracle feedback achieved NDCG@10 of 0.5843 versus 0.2895 for plain BM25. The approach also benefits from optimizing content for findability (documenting why a chunk exists, like coding-agent docs do). The major cost: token usage from repeated exploration that "starts fresh with every new context window."

Model-centric fine-tunes an open-weights model to make the right tool calls for your domain, internalizing what the harness approach learns expensively from scratch each time. A judge labels tool-calling traces as useful or not; the model is then fine-tuned on those traces. Behavior can shift via system prompts ("e-commerce ranking" vs. "financial research diversity"). SID (sid.ai) and Glean's Waldo are early entrants; the approach shows promise but is still in early days.

Turnbull closes by noting converging trends: easier data organization, harness design patterns borrowed from coding agents, and a trajectory where "agents eat the harnesses" — successful patterns get baked into fine-tuned models. The space is a "retrieval primordial goop" whose eventual shape may look nothing like its component parts.