Agentic search models

Traditional search stacks are built from discrete, manually orchestrated components — embeddings, rerankers, BM25, query classifiers — each unaware of the whole pipeline. Agentic search replaces this monolith with an LLM that orchestrates simpler retrieval primitives end-to-end. Frontier models like GPT-5 handle the generic 80% but miss domain-specific nuances (e.g., 'bistro tables' meaning outdoor furniture in a specific store). Purpose-trained agentic search models like SID-1 and Glean's Waldo aim to fill that last 20% by being smaller, faster, and tuned to specific domains and user behaviors. The author predicts a future ecosystem of domain-specific agentic search models — analogous to today's specialized embedding models on HuggingFace — covering e-commerce, job search, and more.