LiveKit Agents v1.8.13 hands off Inference sessions when a gateway drains

1 day ago

LiveKit Agents v1.8.13 hands off LiveKit Inference sessions when a gateway drains, sets a March 1, 2027 removal date for provider pinning, and adds the VLM Run plugin. The Slng plugin is removed, Boson AI is renamed, and Gladia drops its energy filter.

Ask about this presentation

Chapters

  1. 0:00Gateway drain handoff
  2. 0:16Provider pinning
  3. 0:31New providers and models
  4. 0:48Removed and renamed plugins
  5. 1:03Gladia and GPT Live
  6. 1:18Behavior changes
Show transcript
  1. In LiveKit Agents v1.8.13, speech-to-text, text-to-speech and eligible realtime sessions on LiveKit Inference hand off when a gateway drains. Failed WebSocket handshakes now show your quota details.

    Slide text
    LiveKitLiveKit Agents Changed

    LiveKit Agents v1.8.13 hands off Inference sessions when a gateway drains

    https://github.com/livekit/agents/releases/tag/livekit-agents%401.8.13
    Handoff STT, TTS and eligible realtime sessions
    Handshakes Failed WebSocket handshakes show quota details
    GitHub release · livekit-agents@1.8.13 · Behavior changes
  2. If you pin a provider on LiveKit Inference, you now get a warning: explicit pinning will be removed on March 1, 2027. Leave the provider out, and Inference can route your request and fall back.

    Slide text
    LiveKitLiveKit Agents Deprecated

    Pinning a provider on LiveKit Inference now has an end date

    March 1, 2027

    Omit provider to allow routing and fallback

    GitHub release · livekit-agents@1.8.13 · Behavior changes
  3. A new VLM Run plugin adds vision-language chat and System One decisions, including Qwen 3.8. AssemblyAI's realtime agent speech model now runs on LiveKit Inference, and Sarvam realtime speech-to-text adds Saaras v4.

    Slide text
    LiveKitLiveKit Agents Added

    VLM Run brings vision-language models to your agent

    VLM Run New plugin · vision-language chat and System One decisions
    AssemblyAI STT 3.7 Realtime Agent on LiveKit Inference
    Sarvam saaras:v4 on STTRealtime
    GitHub release · livekit-agents@1.8.13 · New providers and models
  4. Two plugins break on upgrade. The Slng speech plugin is removed, so agents built on it need another provider. Boson AI's plugin ships under a new package name, and its avatar plugin stays as it was.

    Slide text
    LiveKitLiveKit Agents Breaking

    Slng is gone, and Boson AI has a new package name

    Slng STT and TTS plugin removed · move to another provider
    Boson AI Now livekit-plugins-boson-ai
    Avatar The Boson AI avatar plugin is unchanged
    GitHub release · livekit-agents@1.8.13 · Breaking changes
  5. Gladia drops its energy filter and forwards all audio, keeping transcript timestamps, and quiet periods no longer end a recording. GPT Live inference refusals now raise a status error instead of a connection error.

    Slide text
    LiveKitLiveKit Agents Breaking

    Gladia now forwards all audio

    Gladia Energy filter removed · transcript timestamps preserved
    Silence Quiet periods no longer end a recording
    GPT Live HTTP refusals raise APIStatusError
    GitHub release · livekit-agents@1.8.13 · Breaking changes
  6. Gemini Live on Inference keeps Gemini's own turn detection, so replies aren't missed, and OpenAI decisions default to GPT-6 Luna. On LiveKit Inference, I would upgrade. On Slng, Boson AI or Gladia, read the release notes first.

    Slide text
    LiveKitLiveKit Agents Changed

    Gemini Live on Inference keeps Gemini’s own turn detection

    Gemini Live Keeps its turn detection to avoid missed replies
    OpenAI DecisionModel() defaults to gpt-6-luna
    Python CPython 3.15 allowed · some plugins capped below it

    LiveKit Agents v1.8.13 release notes on GitHub

    GitHub release · livekit-agents@1.8.13 · Behavior changes

What happened in AI today, in your inbox

One short email each evening. Read it and you’re caught up.

Built with ❤️ in New York