How to Evaluate Live & Voice Agents in ADK- Google Developers Blog

This post introduces native live evaluation capabilities within the ADK (Agent Development Kit). It details how to set up and run evaluations for voice-based agents by simulating user audio input using Gemini TTS. The post provides a concrete example of a multi-agent workflow orchestrated by a `Workflow` object, demonstrating how control and state are passed between agents. It outlines the creation of eval sets with both conversation scenarios (using personas and conversation plans) and fixed conversations. The configuration for live evaluation includes specifying the user simulator type (`llm_audio`), the models for turn-taking logic and audio synthesis, and the criteria for evaluation (e.g., rubric-based trajectory quality). The post also highlights the integration of these live evaluations into the ADK Web interface for interactive debugging, including the ability to play back audio clips of agent responses.