How to Evaluate Live Voice Agents in ADK

Transitioning live voice agents from prototypes to production often stalls on the unpredictability of multi-turn spoken dialogue. Google's Agent Development Kit (ADK) introduces native live evaluation, employing LLM-driven personas that emit synthetic speech via Gemini TTS to stress-test graph-based agent workflows. The system automatically grades audio replies and tool executions against natural-language rubrics, with results inspectable through ADK Web or integrated into automated CI/CD runs.