Skip to main content

Voice Platforms Agent

A managed voice AI platforms comparison between Vapi and Retell. Shows how Waxell Observe traces call lifecycle management, turn-by-turn conversation handling, and platform-specific voice agent configuration across two production voice AI services.

Environment variables

This example requires OPENAI_API_KEY, WAXELL_API_KEY, and WAXELL_API_URL. Use --dry-run to skip real API calls.

Architecture​

Key Code​

Vapi Call Lifecycle​

Three @tool-decorated functions exercise the full Vapi call lifecycle: start, send messages, stop.

@waxell.tool(tool_type="voice")
def vapi_start_call(vapi, assistant_config: dict) -> dict:
    """Start a Vapi call with inline assistant configuration."""
    result = vapi.start(assistant=assistant_config)
    return {"call_id": result.id, "status": result.status}

@waxell.tool(tool_type="voice")
def vapi_send_message(vapi, message: dict) -> dict:
    """Send a message or tool-result through the Vapi call."""
    vapi.send(message=message)
    return {"message_type": message.get("type", "unknown"), "status": "delivered"}

@waxell.tool(tool_type="voice")
def vapi_stop_call(vapi) -> dict:
    """Stop the active Vapi call and get transcript + cost."""
    result = vapi.stop()
    return {"transcript_preview": result["transcript"][:100], "cost": result["cost"], "duration_seconds": result["duration"]}

Retell Agent and Call Management​

Retell exercises agent creation, call creation (phone + web), and call retrieval.

@waxell.tool(tool_type="voice")
def retell_create_agent(retell, config: dict) -> dict:
    result = retell.agent.create(**config)
    return {"agent_id": result.agent_id, "voice_id": result.voice_id}

@waxell.tool(tool_type="voice")
def retell_create_call(retell, agent_id: str, config: dict) -> dict:
    result = retell.call.create(agent_id=agent_id, **config)
    return {"call_id": result.call_id, "call_status": result.call_status}

@waxell.tool(tool_type="voice")
def retell_retrieve_call(retell, call_id: str) -> dict:
    result = retell.call.retrieve(call_id=call_id)
    return {"duration_ms": result.duration_ms, "e2e_latency_ms": result.e2e_latency}

What this demonstrates​

  • Vapi instrumentor -- Vapi.start, Vapi.stop, and Vapi.send traced with tool_type="voice".
  • Retell instrumentor -- CallResource.create/retrieve/create_web_call/create_phone_call and AgentResource.create/retrieve traced.
  • Full call lifecycle -- Vapi: start, send user message, send tool result, stop with transcript and cost.
  • Platform comparison -- @step compares metrics, @decision uses LLM to recommend best platform.
  • Cost and latency tracking -- Vapi call cost and Retell end-to-end latency captured in tool results.

Run it​

# Dry-run mode (no API key needed)
cd dev/waxell-dev
python -m app.demos.voice_platform_agent --dry-run

# Live mode
export OPENAI_API_KEY="sk-..."
python -m app.demos.voice_platform_agent

Source​

dev/waxell-dev/app/demos/voice_platform_agent.py