Small, fast models for the routine calls an agent makes: intent, category, priority, routing, pass or fail. Typed outputs from a fixed set, a confidence score on each, and a fallback to the large model or a person when confidence is low.
Next: Build — Agentic software, built to run.