LiveKit published Python [email protected] on GitHub (2026-09-26; PyPI livekit-agents 1.8.3): Gemini 3.8 Live defaults `tool_behavior=NON_BLOCKING` and sends FunctionTool schemas as `parameters_json_schema` on the text API; OpenAI gains `GPTLiveModel.with_azure` and moves the Realtime inference client into agents core; STT adds AssemblyAI `universal-3-6-pro` plus `language_confidence`, and configurable ElevenLabs realtime chunk duration; voice fixes prevent parallel tools from canceling handoffs, pair tool outputs by call_id, forward frames as-is in fallback/stream adapters, and fix LiveAvatar interruption/plugin registration. Same-day JS `@livekit/[email protected]` deepens agent_turn telemetry and adaptive interruption.

Key Takeaways

  • ✓Shipped: [email protected] / PyPI 1.8.3; JS @livekit/[email protected] same day
  • ✓Gemini 3.8 Live: default NON_BLOCKING tool_behavior; FunctionTool via parameters_json_schema on text API
  • ✓OpenAI: GPTLiveModel.with_azure; Realtime inference client moved into agents core
  • ✓STT: AssemblyAI universal-3-6-pro + language_confidence; configurable ElevenLabs realtime chunk duration
  • ✓Voice correctness: parallel tools no longer cancel handoffs; pair tool outputs by call_id; adapters forward frames as-is
🔗

Project Links & Resources

Direct Access
Direct access to official project resources and documentation
🐙
GitHub Release [email protected]
github.com/livekit/agents/releases/tag/livekit-agents%401.8.3
↗
📖
PyPI livekit-agents 1.8.3
pypi.org/project/livekit-agents/1.8.3/
↗
📖
LiveKit Agents Docs
docs.livekit.io/agents/
↗
📖
LiveKit Realtime Models Docs
docs.livekit.io/agents/models/realtime/
↗
🐙
GitHub Release @livekit/[email protected]
github.com/livekit/agents-js/releases/tag/%40livekit/agents%401.9.1
↗
🐙
livekit/agents repository
github.com/livekit/agents
↗
🔬

In-Depth Technical Analysis

Core Background & Industry Pain Points

Realtime voice agents juggle model tool calls, barge-in, and multi-vendor STT/TTS. Blocking Gemini Live tool behavior stretches audible pauses; Azure GPT Live needs a first-class helper; parallel tools that cancel handoffs drop multi-agent context; adapters that rewrite audio frames drift timbre and timestamps.

Architecture Highlights & Internals

[email protected] (2026-09-26) defaults Gemini 3.8 Live tool_behavior to NON_BLOCKING and sends FunctionTool schemas as parameters_json_schema on the text API; adds GPTLiveModel.with_azure and moves the OpenAI Realtime inference client into agents core. STT: AssemblyAI universal-3-6-pro plus language_confidence; configurable ElevenLabs realtime chunk duration. Voice: parallel tools no longer cancel handoffs; pair tool outputs by call_id; Fallback/Stream adapters forward inner frames as-is; LiveAvatar respects agent.state_updated with interruption/plugin fixes. Same-day JS @livekit/[email protected] deepens agent_turn telemetry and adaptive interruption.

Authoritative Benchmarks & Measured Scores

No public E2E latency/WER numbers—this is an integration/correctness release. Regress Gemini Live non-blocking tool turns, Azure GPT Live connectivity, parallel tool→handoff retention, and AssemblyAI word timings for interruption.

Developer Hands-on Guide

pip install -U livekit-agents==1.8.3, follow Agents docs and Realtime models for Gemini 3.8 Live / Azure GPT Live; align @livekit/[email protected] for JS telemetry. Release: GitHub; package: PyPI.