Google ADK Adds Native Evaluation for Live and Voice AI Agents

Google brings live voice-agent testing into ADK

Google announced native live evaluation for Agent Development Kit (ADK) on August 24, 2026, giving developers a structured way to test voice-based agents across multi-turn spoken conversations instead of relying on manual demos.

The capability lets an evaluation harness drive a live agent with a simulated user whose turns are synthesized as audio, score the agent's spoken responses and tool behavior, and record the resulting trajectory for inspection.

This is a developer-platform capability that is available now in ADK, not a preview of a future model release.

Why live-agent evaluation is different

A voice agent can fail even when its individual answers sound reasonable. It may lose state between turns, call the wrong tool, disclose information before verification, mishandle interruptions or regress after a prompt or model change.

Google's live-evaluation flow treats the whole interaction as the unit of testing.

Developers can define either a conversation scenario, where a simulated user follows a goal and persona while improvising its turns, or a fixed conversation, where user turns are scripted exactly.

The same test cases can run in standard text mode or live mode, which makes it easier to compare logic independently from audio behavior.

Simulated users can speak to the agent

In live mode, ADK can use an llm_audio user simulator. A reasoning model decides what the simulated user should say next, while a separate text-to-speech model synthesizes the audio that is streamed into the live agent.

Google's example uses Gemini models for both roles, including Gemini TTS for generated speech, but the architectural point is broader: the test harness separates turn-taking logic from audio generation.

Personas are prompt-driven, so teams can extend the built-in options to test different communication styles. A max_allowed_invocations setting caps the number of turns to prevent runaway conversations.

Scoring the complete trajectory

ADK supports rubric-based multi-turn evaluation so teams can test requirements that span the whole conversation.

Google's example checks whether a healthcare-style agent confirms identity and validates a date of birth before disclosing appointment information. The evaluation can judge that sequence end to end rather than merely grading each sentence in isolation.

Developers can also attach per-turn metrics to individual responses or tool executions.

This is especially relevant for multi-agent workflows, where control can move between specialized agents while the user experiences one continuous conversation.

CI/CD support turns voice tests into regression tests

The evaluation pipeline can be run from the ADK command line or programmatically through AgentEvaluator.

That means live voice-agent tests can be placed in CI/CD so prompt changes, model updates or tool modifications are checked against repeatable scenarios before deployment.

Google says ADK Web also natively supports live evaluation. After a run, it reconstructs the audio stream into a transcript and shows each turn with transcript text and playable audio for interactive debugging.

What is and is not being released

This announcement is not a new Gemini foundation model. The new capability is the evaluation layer in ADK.

Developers still need API credentials for the live model and the text-to-speech service used by their test configuration. The reliability of an evaluation also depends on scenario design, rubrics, judge-model behavior and the range of simulated users tested.

Synthetic evaluation should therefore complement—not replace—real-user testing, safety review and production monitoring.

Why it matters

Voice agents are becoming more capable at long-running, tool-using conversations, but their evaluation tooling has lagged behind text-agent testing.

Native spoken simulations make it easier to measure whether an agent preserves context, follows procedural rules and calls tools correctly across an entire conversation. Bringing those checks into repeatable automated tests is a practical step toward treating live AI agents like production software rather than demos.