adk-eval-guide

Evaluate ADK-based agents using LLM-as-judge by configuring structured evaluation metrics, defining evalset schemas, and running the evaluation to assess tool trajectories and final response quality. The guide details this complete methodology.

ADK agent evaluations use metrics covering tool trajectory accuracy, final response quality, and safety. Selecting appropriate metrics depends on your specific evaluation criteria and the agent's intended task scope.

Set up scaffolded and non-scaffolded ADK evaluations by following step-by-step running instructions that configure the eval_config and evalsets. Both setups support running evaluations and debugging workflows.

ADK agent evaluations fail or return low scores due to issues in tool trajectories or final responses. Diagnose these common failures by interpreting the LLM-as-judge scores and following the recommended debugging workflows.

Yes, running ADK evaluations requires a structured evalset schema to define test cases and expected outcomes. The guide provides schemas and criteria for configuring your eval_config properly.

Yes, you can tune evaluation criteria for custom ADK agents. The guide includes references for running evaluations, tuning criteria to fit specific use cases, and debugging results.