Thanks — I see you're hitting the libwebrtc warning about "native audio stream queue overflow" when the agent and one user are in the room. That message means audio frames are being produced faster than your consumer code is processing them, so the bounded audio-frame queue fills and older frames are dropped. You can fix this by either increasing the buffer or making your consumer faster: • Increase the queue size (recommended). When creating the audio stream, raise

queue_size_frames

from the small default (usually ~10) to something larger (for example 50):

audio_stream = rtc.AudioStream.from_track(
      track=track,
      sample_rate=48000,
      num_channels=1,
      queue_size_frames=50
  )

• Make your audio processing loop consume frames without blocking (process frames immediately, avoid long synchronous ops inside the loop). • If you accept variable memory usage and latency, set

queue_size_frames=0

to use an unbounded queue (caution: may grow memory). If you want, tell me: • which SDK/language you're using, • your current

queue_size_frames

(if set), • and a short snippet of how you process frames — and I’ll suggest specific values or code changes. Sources: libwebrtc/src/native/audio_stream.rs, with interactive elements