Google's Agent Development Kit (ADK) now supports native live evaluation for voice-based agents, letting developers drive a live agent with a simulated audio user, score multi-turn spoken conversations with rubric-based judging, and run the same eval loop used for text agents. The walkthrough builds a three-stage graph-based live workflow (greeter, DOB verifier, goals agent) running on gemini-live-2.5-flash-native-audio, shows how to author conversation-scenario and fixed-conversation eval cases, configure a user simulator with Gemini TTS voices, run evals via CLI or programmatically through AgentEvaluator for CI/CD, and inspect results with playable audio transcripts in ADK Web.
Questions this post answers
How do I evaluate a live voice agent built with Google ADK?
Google ADK now includes native live evaluation, letting a simulated audio user speak turns to a live voice agent and score the spoken replies within the same eval loop used for text agents. Test cases can be conversation scenarios (simulator improvises turns from a goal and persona) or fixed conversations (scripted turns), run via CLI with adk eval or programmatically through AgentEvaluator for CI/CD. daily.dev surfaces workflow guides like this for teams building CI/CD pipelines around voice agent testing.
What is the user_persona field for in an ADK conversation_scenario eval case?
The user_persona field shapes how the simulated user communicates during a live agent evaluation in Google ADK. For example, the built-in NOVICE persona tells the simulator to share only high-level goals and wait for the agent to ask for specifics, testing how well the agent drives the conversation. Personas are prompt-driven, so custom personas can be added. Developers designing agent test suites can track ADK eval features like this on daily.dev.
How does Google ADK synthesize speech for simulated users in voice agent testing?
Google ADK synthesizes each simulated user turn to speech using the Gemini TTS voice specified in the test_config.json under user_simulator_config, with the llm_audio type. The model field controls the simulated user's turn-taking logic while audio_model handles speech synthesis, and voice_name plus language_code can be adjusted to test agent performance against different voices and accents. daily.dev helps engineers stay current on Gemini TTS and ADK integration details for voice testing setups.
188 Impressions1 Comment