One LLM call runs on every agent trace we ingest. We found a model 30× cheaper that holds quality, verified against real search traffic.