What changed

Google made one evaluation engine generally available in Gemini Enterprise Agent Platform for development experiments and production traces. It includes more than 20 metrics, adaptive rubrics, custom code and model judges, simulators, versioned artifacts, online monitoring and drift alerts.

Why it matters

Using one versioned scoring system before and after launch can expose drift without silently changing the ruler. The release also connects quality scores to complete traces and session logs instead of reducing an agent to one aggregate number.

What to do

Pin a small evaluation set and metric versions, add one deterministic code metric and one adaptive rubric, simulate a slow dependency, then compare local experiments with sampled production traces. Preserve artifacts and manually adjudicate alerting thresholds before using them as release gates.

Caveat

General availability and feature coverage are Google evidence, not independent proof that its model judges are accurate, unbiased or stable. Cloud storage, telemetry, evaluator-model cost, sensitive trace handling and false alerts require tenant-specific testing.