LiveKit guide: building a production-ready AI voice agent
LiveKit published a guide on taking AI voice agents from demo to production, organized around three questions: whether the agent sounds smart and natural, whether teams can tell how it performs in production, and whether it scales reliably and cost-effectively. Among its recommendations, LiveKit advises measuring the time from pickup to the agent's first word as a separate metric — one unnamed company found about three of every four calls ended at the greeting step — and passing only greeting-level context with the call while loading the rest in the background. It also recommends LiveKit's turn detector (on by default in AgentSession) with adaptive interruption handling, asynchronous tool calls with filler phrases to avoid dead air, noise-cancellation testing, and expressive TTS; for latency it cites LiveKit benchmarks showing Gemma 4 31B at 192 ms time to first token versus 966 ms for GPT-5.5, 1,006 ms for GPT-4.1 and 1,095 ms for Gemini 3.0 Flash, and says some clients moving from GPT-4.1 to Gemma saved 300–500 ms at p50.
Most teams can demo a voice agent quickly; the harder, less documented work is the latency, turn-taking and observability tuning that decides whether it survives real callers.