The Core Update

Building live voice agents is tricky. Demos often hide real-world issues. Google's ADK now offers native evaluation for these agents. You can simulate audio-based conversations. Then you score agent replies. This all happens within your existing ADK evaluation loops. The goal is confident deployment. You get measured, trusted agent performance.

Official Source: Google Announcement

Technical Impact & Mechanism

Previous evaluation primarily focused on text. That missed voice nuances. Agent behavior changes easily. Tools misfire. Context drops. Interjections get ignored. This update addresses those points. ADK's new system uses a simulated user. This user speaks turns as actual audio. The agent's spoken responses are then evaluated. This validates speech, timing, and recovery.

Under the hood, you define evaluation cases. These use a JSON structure. conversationscenario is a key type. Here, you set a goal and a user persona. The simulator improvises the conversation. Personas like NOVICE guide the simulator. They ensure varied user behavior. Max turns (maxallowed_invocations) prevent runaway tests. Your agents can be chained too. Use ADK's Workflow class. It orchestrates multiple Agent instances. Each agent operates with models like gemini-live-2.5-flash-native-audio. Workflow carries session state. Conversation history persists across agent handoffs. The audio stream stays open throughout. This mirrors a single, continuous user interaction.

CONSOLE // JSON SYNTAX_CHECK: OK 
{
  "eval_id": "example_voice_scenario",
  "conversation_scenario": {
    "starting_prompt": "Hey there.",
    "conversation_plan": "You need to confirm your identity. Give your DOB as 1985-07-12 when asked. Listen to appointment details. Ask one follow-up question. Then end the call.",
    "user_persona": "NOVICE"
  },
  "session_input": {
    "app_name": "your_live_workflow",
    "user_id": "test_user_001",
    "state": {}
  }
}
This JSON defines a test. The simulator acts out the conversation_plan. It uses the NOVICE persona.

Action Plan for Developers & Businesses

  1. Update ADK Dependencies: Ensure your ADK SDK and related libraries are current. Access the new evaluation APIs.
  2. Model Agent Workflows: Design your live voice agent interactions. Use ADK's Workflow to sequence Agent instances. Ensure state and context pass correctly.
  3. Author Voice Eval Cases: Create JSON conversation_scenario files. Describe user goals and personas. Simulate diverse user interactions.
  4. Automate Evaluation in CI/CD: Integrate these new audio evaluation loops. Run them as part of your regular testing. Catch regressions early, before production.
Need help optimizing your digital systems or building robust AI agents? Check out my Case Studies & Work to see how I tackle complex challenges. Ready to build something reliable? Contact me for a direct discussion.
WR
AUTHOR // SOLUTIONS ARCHITECT

MUHAMMAD WALEED RAZA

I am a Solutions Architect and Digital Growth Engineer specializing in server-side tracking, CRM integrations, and conversion data recovery. I build deterministic, data-driven systems for high-growth brands and enterprise agencies worldwide.

HIRE ON UPWORK START A PROJECT