What we covered
A voice agent is judged on one thing before anything else: whether the pause before it answers feels human. That constraint decides the architecture, so the session works backwards from it.
- Why voice agents became viable now, and what changed in the model layer to make them so.
- The two stacks side by side: an STT → LLM → TTS pipeline versus native speech-to-speech, and what each one costs you.
- Where LiveKit's real-time audio transport and the Gemini Live API fit in that picture.
- Building the agent itself: tool definitions, turn detection, and handling a caller who interrupts.
- The latency budget, broken down hop by hop, and how to decide which hop gets the milliseconds.
- Agentic memory, and how to give an agent context without paying for it in response time.
- What production adds: deployment, observability, and the failure modes that only show up on a real phone line.
Format
A technical talk with a live demo of a working voice agent, aimed at engineers who will be writing this code rather than commissioning it.
Who it's for
Engineers and architects building conversational products, particularly anyone who has a working prototype that feels a beat too slow and needs to know where the time is going.