What problem does it solve? Building an AI agent is only half the battle — knowing whether it actually works reliably is the hard part. This Skill guides you through the Agent Platform evaluation methodology so you can measure agent quality with structured datasets, built-in and custom metrics, and iterative failure analysis instead of guessing. ## Core Features & Use Cases - Quality Flywheel Workflow: Follow the five-stage loop (prepare data, run inference, grade traces, analyze failures, optimize) using agents-cli eval commands like generate, grade, compare, analyze, and optimize. - Metrics Selection & Custom Metrics: Choose from built-in metrics such as multi_turn_task_success, multi_turn_tool_use_quality, hallucination, and safety, or define custom LLM-as-judge and code-execution metrics in eval_config.yaml. - Dataset & Trace Schemas: Author single-turn, multi-turn, and multi-agent evaluation datasets, synthesize conversations with user simulation, and handle special cases like google_search built-in tools and multimodal inputs. - Use Case: Your agent's tool_use_quality score is low. Open the grade results, inspect the rubric verdicts, refine the tool descriptions, rerun generate and grade, then use eval compare to confirm the fix improved the metric without regressions. ## Quick Start Ask the assistant to run an evaluation of your ADK agent and analyze any failing metrics using the agents-cli eval workflow.
