What key metrics would you track to evaluate the safety and quality of a conversational AI agent at Google DeepMind?