AI agents should leverage expert-level search capabilities rather than being constrained to the simplified interfaces designed for casual human users. The evolution of retrieval for LLMs has moved from basic vector search to hybrid approaches, and is now entering a third stage where agents can issue precise, complex queries — filtering by metadata, grouping results, choosing ranking strategies — much like a professional quant doing financial analysis. Perplexity's 'search as code' announcement signals this shift going mainstream. In practice, enabling this requires giving agents a description of available fields and ranking profiles (e.g., via YQL in Vespa), allowing them to construct targeted queries rather than relying on vague keyword matching. Developers building agentic search should focus on providing a rich toolbox of capabilities rather than optimizing for the lowest-common-denominator human query experience.
