Our take
A guidance product is a designed path through the moments that decide it, not one clever prompt. The states a user arrives in, the move that fits each, and the transitions between them are an asset a general model only improvises. Map them, and behavior becomes something you can build and test per stage.
When talk is not enough
Sometimes the agent has to act, not only advise, so you give it tools that turn talk into an outcome. Every tool is new power and a new door in, so you start with the simplest workflow that works and add autonomy only where it pays. The right answer is sometimes not to build an agent at all.
Grade the whole path
Once an agent takes steps with tools, grading the final answer hides where it went wrong and how often it will. You evaluate the trajectory and the reliability across repeated runs, not a single lucky pass. And in voice, a lot of the character lives in timing, so the harness reaches into turn-taking and latency, not just the words.