Google says agent and model evaluations in Gemini Enterprise Agent Platform are now generally available, giving teams a unified way to test agents during development and grade deployed behavior after launch.

The service is aimed at a practical problem in agent deployment: teams need to know whether an agent can finish a task, call tools correctly, follow policy, and keep quality stable after it is exposed to real users. Google says the GA release supports local experiments for fast iteration, plus grading of existing sessions and traces when an agent is already deployed with telemetry enabled.

What is included

Google's announcement says the evaluation service includes more than 20 pre-built metrics, DeepMind-backed adaptive rubrics, custom code-based metrics, and LLM-as-a-judge metrics stored in a centralized, versioned registry. The company also describes metrics for task success, tool use quality, and safety, including checks around tool selection, argument correctness, schema compliance, and policy-sensitive responses.

The platform can also generate test cases through a case generator, user simulator, and environment simulator. That matters because hand-written eval sets often miss edge cases, especially for agents that depend on external tools or multi-step workflows.

Why it matters

Agent products are moving from demos into systems that need release gates and operational monitoring. Google's GA release does not prove that automated evals can fully replace human review, but it gives Gemini Enterprise Agent Platform customers a more formal loop for measuring agent behavior before and after deployment.

For developers, the most important shift is that evaluation is becoming part of the agent platform itself, rather than a separate notebook or after-the-fact audit.