Unified Evaluation Engine Hits Gemini Enterprise Platform

On July 31, 2026, Google Cloud moved Agent and Model Evaluations to general availability within the Gemini Enterprise Agent Platform. The release provides a unified evaluation engine designed to measure AI agent quality consistently across the entire lifecycle. If you build and deploy agentic systems, this allows you to apply the identical metrics and rubrics to local code experiments and live production traffic.

The service addresses the measurement drift that occurs when teams use different frameworks for development testing and production monitoring. Developers can access the unified engine through the Agent Platform SDK, the agents-cli tool for local workflows, or the Google Cloud Console for visualizing score-over-time charts and drift alerts.

Core Evaluation Capabilities

The platform introduces a centralized Metric Registry for managing organization-wide evaluation standards. This repository supports versioning and handles three distinct metric types to evaluate and test AI agents effectively.

Metric Type	Implementation	Primary Use Case
Pre-built Metrics	Over 20 managed metrics	Measuring quality, safety, groundedness, and tool-use accuracy
Custom LLM Metrics	Natural language criteria	Evaluating complex semantic responses using a judge model
Custom Code Metrics	Python functions	Programmatically validating agent outputs and formatting

To improve the reliability of the custom LLM metrics, Google collaborated with Google DeepMind to develop Adaptive Rubrics. These rubrics dynamically tailor judging criteria to each specific input case. This approach avoids the common brittleness associated with static LLM-as-a-judge prompts.

Simulation and Testing Tools

Testing autonomous systems requires realistic environment interactions before production deployment. The release includes built-in simulation tools to automate this process.

The User Simulator executes multi-turn conversation cases without requiring manual scripting for each turn. The Environment Simulator emulates backend systems, allowing developers to test how agents handle failing or slow APIs. This enables robustness testing without generating load on live infrastructure or requiring complex mock servers.

Platform Context and Models

This evaluation release follows a broader push to mature the Gemini Enterprise Agent Platform. On July 29, Google moved Agent Memory Bank for long-term context, Agent Runtime for tasks lasting up to seven days, and Agent Identity for auditable access control to general availability.

The platform supports Gemini 3.6 Flash, which provides improved planning capabilities and lower price points for high-volume tasks. It also supports third-party routing, including Anthropic’s Claude Opus, Sonnet, and Haiku models.

Enterprise security and compliance teams are shifting focus toward standardizing evaluation and governance tools for AI agents. Early adopters like Commerzbank AG and Liberty Global are utilizing these evaluation pipelines to enforce compliance policies, ensuring that only formally vetted agent configurations reach production environments. Organizations adopting this platform should integrate the metric registry directly into their CI/CD pipelines using the Agent Platform SDK to catch regressions before deployment.