Evaluate ADK agent performance using the Agent Platform Generative AI Evaluation Service

Why it matters

Agent evaluation needs evidence at several layers. A single end score can hide a wrong tool, a bad trajectory, or a plausible final answer produced for the wrong reason.

What to practise

Prepare an evaluation dataset and measure tool choice, action trajectories, response quality, and a custom criterion. Inspect the results instead of treating the aggregate as self-explanatory.

How it fits

This lab applies every Bite in subsection 4.1: evaluation sets, layers, success criteria, metrics, autoraters, and regression gates.