Gemini Enterprise Agent Platformのエージェント・モデル評価機能がGAにAgent and Model Evaluations in Gemini Enterprise Agent Platform are now GA

Google has moved the evaluation service in its Gemini Enterprise Agent Platform to general availability, giving developers a formally supported way to measure how well their AI agents behave. The change matters because judging agent quality is one of the harder problems in shipping production generative AI: outputs are probabilistic, behavior can shift between a controlled test and real-world use, and teams often lack a consistent yardstick to know whether a change made things better or worse.

At the center of the release is what Google describes as a unified engine for measuring agent quality across two very different settings: local development experiments and live production traffic. That consistency is the main selling point. Agents frequently look reliable in a notebook or a small test harness, then behave differently once real users, unpredictable inputs, and edge cases arrive. Running the same evaluation framework in both phases lets developers compare results on equal footing instead of maintaining separate, incompatible tooling for offline testing and online monitoring.

The service ships with more than 20 pre-built metrics, some of which Google says draw on DeepMind research. Pre-built metrics generally lower the barrier to getting started, since teams can begin scoring agents without designing their own rubrics from scratch. Evaluation suites for agents typically span several dimensions — the quality and correctness of a final answer, whether the agent followed a sensible sequence of steps, how it used external tools, and whether responses stay grounded in provided context. Google has not framed this as a new model, but as infrastructure for assessing the agents developers build.

Many modern evaluation metrics for open-ended tasks rely on a technique often called "LLM-as-a-judge," in which a capable model grades another system's outputs against defined criteria. This approach scales better than manual human review but introduces its own considerations around bias and reliability, which is part of why vendor-provided, research-backed metrics are attractive: they offer a common baseline that is easier to reason about and reproduce. The DeepMind connection is notable here, since Google DeepMind — formed in 2023 from the merger of DeepMind and Google Brain — is the research organization behind the Gemini models themselves.

The evaluation service fits into a broader Google stack for building and running agents. The presence of Vertex AI among the associated tags points to integration with Google's managed machine learning platform, which already offers components for deploying and operating agents at scale. Positioning evaluation as a shared engine, rather than a standalone tool, reflects a wider industry pattern in which measurement, observability, and quality control are increasingly treated as core parts of the agent lifecycle rather than an afterthought.

That pattern is visible across the market. Competing and complementary offerings — such as LangChain's LangSmith, OpenA