Turn design, error recovery, handoff to a human, and what the interface does while the model is thinking. I rebuilt SkillsAware's chat as a multi-agent model with one agent per workflow, because a single assistant trying to do everything does all of it slightly wrong.
