Thanks — I’m responding to your question about how to avoid cold starts in a deployed LiveKit agent. Summary — what you can do (high level) • Use a paid LiveKit Cloud plan with always-on agents (Ship or higher) so worker instances remain warm. • Enable cold-start prevention in your deployment configuration to keep workers pre-warmed. • Prewarm models and resources during agent initialization so they’re loaded before the first request. • Pre-download large model files in your Docker build so containers start ready to serve. Concrete steps you can take 1. Choose a paid LiveKit Cloud plan ◦ Deploy on a paid plan (Ship or higher). Agents on those plans are kept warm and ready for immediate connections; Build-tier deployments may experience cold starts. 1. Enable cold start prevention ◦ Turn on cold start prevention in your deployment configuration so the platform keeps the needed workers pre-warmed for deterministic startup latency. 1. Prewarm model and library initialization at worker startup ◦ Add a prewarm function that loads models into worker memory during process startup instead of on first request. Example pattern in Python:

def prewarm(proc: JobProcess):
         proc.userdata["vad"] = silero.VAD.load()
     server.setup_fnc = prewarm

• Prewarming ensures shared objects (VAD, STT, TTS, LLM clients) are ready and reused across sessions, avoiding per-connection model load cost. 1. Pre-download model files during container build ◦ Add a step in your Dockerfile to download large model files at build time:

RUN uv run src/agent.py download-files

• This prevents runtime network/model-download latency when containers start. 1. Right-size and monitor your deployment ◦ Ensure your deployment has sufficient capacity so LiveKit Cloud can keep the workers warm. Monitor warm-worker availability and startup latency and adjust sizing if needed. If you want, I can: • Walk through where to enable cold-start prevention in your current deployment config. • Review your agent startup code (or Dockerfile) and point out exactly where to add prewarm/download steps. • Suggest what to prewarm first (e.g., VAD, STT) based on which features you use. Which would you like help with next? If you share your deployment type (LiveKit Cloud plan) and a short snippet of your agent startup or Dockerfile, I’ll give targeted instructions. Sources: How to match the latency of the homepage agent | LiveKit | How much can LiveKit scale? | LiveKit | Collect recording consent with tasks | LiveKit Documentation | Silero VAD plugin | LiveKit Documentation | VAD Metrics | LiveKit Documentation