An analysis comparing Meta Muse, OpenAI Dots and Uber’s driver assistant argues that the central challenge for agents that speak first is deciding when to intervene rather than generating a response. It proposes weighing impact, the likelihood that the user will act, how long the opportunity remains relevant and whose interests are served to choose among immediate delivery, bundled notifications, deferral and discarding. Predictive models or judgment models that return probabilities could filter candidates before text generation, but learning individual responses requires separate data. This is a design proposal, not an experiment establishing equivalent effects across the three services.