Conversational Voice AgentsConversationalVoiceAgents

Conversational Voice Agents is a production-oriented platform for configuring and running voice agents across more than one real-time engine. The control plane owns tenant configuration and call state; stateless workers execute calls from a shared, versioned runtime contract.

The implementation is maintained in private repositories. This project page exposes the architecture and engineering decisions without publishing private source code, credentials, or operational configuration.

Voice-agent runtimes evolve quickly. Tying product configuration, business logic, and observability directly to one engine makes migration expensive. This platform separates the control plane contract from the worker implementation so LiveKit and Pipecat can execute the same agent profile.

Every call is described by a typed runtime profile: providers and models, prompts, tools, transport, recording, voicemail behavior, idle handling, duration limits, transfers, and reporting. Unknown providers fail explicitly instead of silently falling back.

The LiveKit worker and Pipecat worker share generated wire contracts, not engine internals. Each runtime uses its native pipeline and lifecycle while emitting the same signed events back to the control plane.

The system supports inbound, outbound, and web calls with transport-specific routing. It models answering-machine detection, call recording, cold transfers, warm-transfer boundaries, and the provider infrastructure required for each path.

Lifecycle events are idempotent and HMAC-signed. Metrics include model usage, latency, transcripts, end reasons, and failure context. OpenTelemetry spans make the real-time pipeline inspectable without coupling the control plane to a single worker.

The architecture turns a voice-agent engine into a replaceable execution layer. Product behavior remains portable, migrations can happen incrementally, and new runtimes can be evaluated against the same contract and test scenarios.