ADKでライブ・音声エージェントを評価する方法How to Evaluate Live & Voice Agents in ADK

Google has added native live evaluation to its Agent Development Kit (ADK), giving developers a built-in way to test voice and other real-time agents automatically. According to a post on the Google Developers blog, the feature targets teams trying to move live voice agents from working demos into reliable production systems, where the unpredictability of genuine multi-turn conversations makes manual testing impractical.

The core problem the update addresses is that live voice interactions are hard to test with traditional, script-based methods. A conversation can branch in many directions, users interrupt, rephrase, and change their goals mid-sentence, and the agent's response at each turn depends on everything that came before. Checking that behavior by hand does not scale, and static test scripts tend to miss the edge cases that surface once real users are involved.

ADK's answer is to test graph-based agent workflows against LLM-driven simulated users. Rather than replaying fixed transcripts, the framework uses a language model to play the role of a user, generating conversational turns dynamically and reacting to what the agent says. This lets a single test scenario explore many possible dialogue paths and probe how the agent handles follow-up questions and unexpected inputs. Because the evaluation is native to ADK, developers can define these tests alongside the agents themselves rather than assembling a separate testing stack.

ADK, which Google introduced as an open-source framework for building and orchestrating AI agents, models agent behavior as workflows that can include multiple steps, tools, and decision points, the "graph-based" structure referenced in the announcement. It is designed to work with Gemini models and integrates with Google's broader agent tooling, including Vertex AI and the Agent Engine runtime, though the capability described here focuses specifically on the live and voice case.

The approach reflects a wider industry pattern often called "LLM-as-a-judge," in which one model is used to generate or assess the behavior of another. Applying that idea to simulated users is a natural extension: an LLM can improvise plausible customer questions, complaints, or clarifications, providing broader coverage than a human tester could write by hand. The trade-off, which teams typically weigh, is that model-generated evaluations can carry their own biases and inconsistencies, so simulated testing is generally treated as a complement to, rather than a replacement for, human review and real-world monitoring.

Voice agents in particular raise challenges that text-only agents do not. Real-time audio introduces latency, turn-taking and interruption behavior, and errors from speech recognition, any of which can derail a conversation. Google's Gemini Live API and comparable real-time interfaces from other providers have made building such agents more accessible, but the tooling for rigorously evaluating them has lagged behind. Native live evaluation in ADK appears intended to help close that gap.

For developers, the practical value is the ability to run automated regression tests as an agent evolves. When a prompt, tool, or model changes, a suite of simulated conversations can be replayed to check whether the agent still handles key scenarios correctly, reducing the risk that an improvement in one area quietly breaks another. This is the same discipline that automated testing brought to conventional software, adapted to the non-deterministic nature of conversational AI.

The feature sits within a competitive landscape where agent evaluation is a growing focus. Companies building agent frameworks and observability tools are increasingly offering tracing, scoring, and simulation features, as production deployments expose how brittle early agents can be. By folding live evaluation into the kit used to build the agents, Google is positioning testing as a first-class part of the development loop rather than an afterthought.

As with any evaluation methodology, results are likely to depend heavily on how the simulated users and success criteria are configured. Teams will still need to define what a good conversation looks like for their use case and interpret the outcomes accordingly. Even so, the addition gives ADK users a documented, native path for stress-testing live and voice agents before they reach real users, an important step for anyone weighing the move from demo to production. The blog post is framed as a tutorial, suggesting Google intends to walk developers through configuring and running these evaluations in practice.