Agent and Model Evaluations in Gemini Enterprise Agent Platform are now GA

Agent and Model Evaluations in Gemini Enterprise Agent Platform are now generally available, according to the Google Developers Blog. The Agent Platform evaluation service is out of limited release and positioned as a unified engine for measuring agent quality. That same engine is meant to cover both local development experiments and live production traffic, so quality signals are not split across separate tooling stacks.

On the technical side, the service ships more than 20 pre-built metrics plus DeepMind-backed adaptive rubrics. Teams can also define custom code-based metrics and LLM-as-a-judge metrics, and store them in a centralized, versioned registry. Integration is direct: the Agent Platform SDK, agents-cli, and ADK wire the evaluation path into existing agent workflows instead of requiring a parallel evaluation harness.

For engineers and builders, the practical shift is consistency. An agent scored during local experiments can be measured with the same metric definitions and registry entries once it serves production traffic. Versioned metrics reduce the risk that a rubric changes silently between experiment and release, and custom judges let teams encode domain-specific quality rules without abandoning the shared platform.

In market terms, agent quality work often fragments: offline notebooks, ad hoc scripts, and production monitors use different definitions of “good.” A GA evaluation service that spans local and live traffic, with a shared registry for pre-built, adaptive, code-based, and LLM-as-a-judge metrics, is Google’s bet that enterprise agent teams want one measurement layer rather than bolting eval onto each environment.

The immediate takeaway is to treat evaluation as first-class platform infrastructure: pick or author metrics in the registry, run them through the SDK or agents-cli during development, and re-run the same definitions against production. Watch how adaptive rubrics and custom judges behave on real traffic versus experiment sets, and whether the centralized registry becomes the single source of truth for agent quality across teams.