RenderHazal Mestci4 min readintermediate
Search for agents splits discovery from proof
Summary
Exa’s Agent splits search for AI agents into a discovery phase (candidate generation) and a verification phase (evidence check). The design uses a powerful model for generation and cheaper models for parallel verification, optimizing latency, token usage, and cost. It treats the web as a structured database rather than a ranked list, and runs the verification sub‑agents as independent, fault‑tole…
- Agent‑centric search requires low latency (200‑300 ms) and token‑efficient retrieval, unlike human‑focused search.
- Treating the web as a database means semantic search and structured extraction outweigh simple keyword ranking.
- Exa’s architecture separates discovery (expensive, high‑capacity model) from verification (cheap, narrow model) to control cost and improve quality.
- Verification sub‑agents are independent, enabling massive parallelism and easy orchestration via a workflow engine.
As AI agents become the primary consumers of web information, search systems must shift from delivering ranked link lists to providing structured, verified data. Exa’s split‑architecture demonstrates a practical way to meet the latency, cost, and accuracy constraints of agent workloads, and shows h…
6/10
