Speech output as agent tools on the platform TextToSpeech engine: speak reads text aloud (and returns when it has been spoken, so the agent can sequence it) and stop_speaking interrupts it. Nothing leaves the device unless the user's TTS engine is a network one.