Thanks — I understand you're seeing agents randomly dropped from calls even though your Python worker logs show no errors and there are no resource constraints. I can help narrow this down. A few quick clarifying questions that will point us to the most likely causes: • Are you running multiple agent instances (or could other developers be) using the same agent name or credentials? Shared/dev instances can cause "crosstalk" where another instance accepts a session. • Do drops correlate with deployments, autoscaling, or container restarts (even rolling updates)? Graceful drain / rollout mismatches can terminate sessions. • Can you check the client/SDK-side logs (browser or mobile) and WebRTC connection state events around a drop? Do you see ICE failures, signaling disconnects, or reconnection attempts? • Are you seeing any network instability (packet loss, NAT timeouts, mobile network switching) on the agent side around the time of the drops? • How frequent are the drops (e.g., once per hour, daily, only during certain calls) and do they affect all agents or a subset? What I can do next once you share answers or logs: • If crosstalk is possible, I’ll suggest ways to prevent duplicate agent instances from stealing sessions. • If deployments/rollouts look related, I’ll walk through drain_timeout and rollout considerations. • If SDK/WebRTC events show ICE or signaling issues, I’ll guide you on what specific connection metrics or logs to capture (ICE state, pc connection state, and reconnection attempts). • If signal handling or wrapper scripts are used for the agent process, I’ll point out how to ensure signals are forwarded so the agent can shut down gracefully. Reply with the answers to the questions above and any relevant timestamps or SDK logs around a recent drop and I’ll help you diagnose the next steps. Sources: Deployment reliability on LiveKit Cloud | LiveKit | Avoiding crosstalk when multiple developers share a LiveKit project | LiveK... | Where to find your Agent Logs | LiveKit, with interactive elements