with self hosted agents on LK cloud, is there a way to have a metric of avg agent sessions per worker?

Thanks for reaching out, Sebastián. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.

1. a worker process 2. pods 3. LiveKit Cloud with self‑hosted agents 4. no 5. 24h average

Thanks — this is helpful. I’ll walk you through how to get a 24‑hour average of agent sessions per worker (pods) for LiveKit Cloud with self‑hosted agents. Summary of what you need • Export per‑worker session metrics to Prometheus (one metric that increments on session start or reports current active sessions per worker). • Export worker presence so Prometheus can count pods (e.g.,

up{job="livekit-agent"}

or a

livekit_worker_info

metric with

worker_id

label). • Use PromQL to compute the 24h average sessions per worker. If you already have no metrics scraping (you answered “no”), next step is to enable metrics on the agent pods: • Ensure the agent process exposes Prometheus metrics (HTTP /metrics). • Add a Prometheus scrape config for the agent pod endpoints (or use ServiceMonitor if using Prometheus Operator). Recommended metrics to expose (or create) • A counter

livekit_session_started_total{worker_id="<id>"}

that increments when a session starts (or) • A gauge

livekit_worker_active_sessions{worker_id="<id>"}

reporting current active sessions per worker. Example PromQL (24‑hour avg) 1. If you have a session start counter: ◦ Sessions started per worker over 24h: sum(increase(livekit_session_started_total[24h])) by (worker_id) • Average sessions per worker (24h): avg(sum(increase(livekit_session_started_total[24h])) by (worker_id)) 1. If you have an active‑sessions gauge and want time‑averaged active sessions: ◦ Per worker 24h average: avg_over_time(livekit_worker_active_sessions[24h]) by (worker_id) • Overall average across workers: avg(avg_over_time(livekit_worker_active_sessions[24h]) ) If you need a ratio of total sessions to worker count (workers may come and go): • Total sessions started over 24h: sum(increase(livekit_session_started_total[24h])) • Worker count (instant or average over 24h): count(up{job="livekit-agent"}) # instant avg_over_time(count_over_time(up{job="livekit-agent"}[5m])[24h]) # approximate • 24h average sessions per worker: sum(increase(livekit_session_started_total[24h])) / count(up{job="livekit-agent"}) How I can help next • If you tell me which metric names your agent exposes (or paste a sample /metrics output), I’ll give the exact PromQL/Grafana panels. • I can provide a scrape_config snippet for Prometheus or a ServiceMonitor for Prometheus Operator if you’re on Kubernetes. Sources: Voice Agent Architecture: STT, LLM, and TTS Pipelines Explained | LiveKit, with interactive elements

what's the difference between session.start and room.connect in the LK agent?

Thanks — good question. Here’s the short difference I’m responding to: "what's the difference between

session.start

and

room.connect

in the LK agent?" •

room.connect

◦ Low‑level LiveKit SDK call that joins a participant to a LiveKit room (handles signaling/WebRTC, auth with a token, and media transport). ◦ Use it when you only need to establish the WebRTC connection for a client or participant. •

session.start

◦ High‑level Agents framework call that starts an agent session. It internally performs the room connection (via

room.connect

or equivalent) and also initializes the agent lifecycle: STT, LLM, TTS, VAD, I/O plumbing, and any agent orchestration logic. ◦ Use it when you want the full agent pipeline and session management rather than just a raw room join. If you want, tell me which SDK/class you’re using (AgentSession vs Session vs plain Room) and I’ll point to the exact code paths or events that fire for each. Sources: Agent session | LiveKit Documentation | How to Build a Background Observer for Voice AI Guardrails | LiveKit | Build agent experiences faster with Session APIs | LiveKit, with interactive elements