Some LLM calls in an agent can be skipped: Jev returns a label, not text

Jev by TypeSafe AI writes nothing: a prompt goes in, a label, score, or route comes out.

If routing moves to a layer like this, the large model will be invoked only where text is needed.