AI News vs LiveKit Agents

livekit-agents is the orchestration layer for realtime voice AI agents, now deep in a high-cadence 1.6.x line shipping every few days. The work splits cleanly in two: a sprawling provider-plugin surface (OpenAI Realtime, Google Gemini 3, ElevenLabs, AssemblyAI, Deepgram, Cartesia, Fishaudio, Phonic, Soniox, xAI, Azure, Krisp) and the harder core primitives — turn detection, interruption handling, and tool-call correctness across agent handoffs. The recent directional releases signal LiveKit betting on the orchestration problems, not just provider coverage.

The arc points toward being the neutral conversation engine that sits above any speech or LLM vendor, where the defensible value is turn-taking, latency, and tool-execution semantics rather than any single model. Provider breadth keeps widening release over release, but the marquee features — async tools, the turn detector — are all about how the agent behaves mid-conversation. Expect the core voice loop to keep hardening while the plugin roster grows.

Next releases likely continue the two-track pattern: rapid provider model and parameter updates alongside deeper investment in the conversation loop — more work on interruption, latency, and long-running tool behavior building on the async-tools and turn-detector foundations.