LoginSign Up
← Back to all sparks
L

LiveKit Agents

AI-ASSISTANTS
Velocity3.9

Framework for building real-time voice and multimodal AI agents over WebRTC.

docs.livekit.io/agents ↗

LiveKit Agents added Answering Machine Detection — voice agents are becoming a serious telephony runtime.

answering-machine-detectiontelephonyvoice-agentsprovider-pluginsbarge-inmcp-deprecation
◆Current state
LiveKit Agents is releasing roughly twice a week along the 1.5.x line, accumulating telephony-grade primitives around its voice loop. The headline is Answering Machine Detection in 1.5.9 — an LLM-classified detector for what kind of endpoint an outbound call hit. Surrounding work is split between reliability (barge-in cooldown, interruption guards, preemptive-generation tuning, observability retries) and provider breadth (Perplexity Responses, Soniox, Speechmatics, Cerebras, xAI, Rime WebSocket TTS). The mcp_servers parameter was also deprecated on Agent and AgentSession.
◆Where it's heading
The product is converging on a real contact-center runtime, not just a realtime meeting agent. AMD, warm transfer, DTMF handling, recording retries, and avatar join/playback metrics are the feature surface phone deployments demand. The provider plugin universe keeps widening; LiveKit positions itself as the neutral broker between voice models and the actual network. Internal cleanups (mcp_servers deprecation, instruction parts, AvatarSession base class) suggest a tidying pass before a 1.6 cut.
◆Prediction
Expect more telephony primitives — supervisor barge-in, richer DTMF flows, call-recording controls — and a unified MCP configuration surface across Agent and Session as the mcp_servers deprecation lands fully.

◆Recent moves

  1. 13h ago

    Agents 1.5.12: deprecate mcp_servers; Perplexity Responses; AvatarMetrics

    1.5.12 deprecates mcp_servers on Agent and AgentSession, adds Perplexity Responses LLM, gpt-realtime-whisper STT, multi-message generation in realtime, UserTurnLimitOptions for cutting off long user speech, and AvatarMetrics for join/playback latency. Provider breadth plus API-surface cleanup — fits the contact-center hardening arc.

    View source ↗
  2. 1d ago

    Agents 1.5.11: routine version bump

    1.5.11 is a bot-cut version bump with no published release notes. Routine release in the high-cadence 1.5.x line.

    View source ↗
  3. 3d ago

    Agents 1.5.10: live LLM swaps and Speechmatics STT in inference

    1.5.10 adds inference.LLM.update_options for live model swaps, Speechmatics STT to inference, Rime Coda model, and shutdown-callback handling. The live-swap mechanism is the standout — it lets running sessions change model without tearing down the agent.

    View source ↗
  4. 7d ago

    Agents 1.5.9: Answering Machine Detection for outbound calls

    ⚡ SPARK

    1.5.9 introduces Answering Machine Detection — the feature that turns LiveKit Agents into a credible outbound-calling platform, not just a realtime meeting agent. Bundled with WarmTransferTask additions, Perplexity LLM, and Rime WebSocket TTS.

    View source ↗
  5. 15d ago

    Agents 1.5.8: barge-in cooldown and observability retry

    1.5.8 adds barge-in cooldown for corrections, switches Fish Audio TTS to WebSocket for faster inference, retries session-recording upload, and adds Soniox TTS. The barge-in cooldown is the meaningful one — handling the case where a user corrects themselves without re-triggering the interruption pipeline.

    View source ↗
  6. 20d ago

    Agents 1.5.7: dynamic endpointing, playback metric, Runway and Cerebras plugins

    1.5.7 is a heavy-throughput release: dynamic endpointing alpha param, playback_latency metric, Runway Characters avatar plugin, Cerebras LLM plugin, expanded MCP header updates, double-encoded JSON tool-arg handling. Lots of provider plumbing rather than a single headline, consistent with the breadth-over-depth phase.

    View source ↗