Concept 8 · Google · AI Community 2026Agent Evaluations (adk eval)
04a, 04b & 07a–c

.evalset.json: tests written in JSON format containing multi-turn conversations.

eval_config.json: grades every multi-turn conversation of the .evalset.json on two axes:

tool_trajectory_avg_score (1.0): verifies the exact sequence of tools and sub-agent transfers executed.

response_match_score (0.5): measures semantic/ROUGE similarity against golden reference responses.

--print_detailed_results: to have a comparison table.

Web UI evals: you can create evals from the Web UI, but you will probably have to revisit them (user_id example in Registratore's tool).

eval_config.json & Terminal
{
"criteria": {
"tool_trajectory_avg_score": 1.0,
"response_match_score": 0.45
}
}
# Run automated evaluation suite:
$ uv run adk eval adk_agent_app \
adk_agent_app/evals/eval_set_4_full_workflow_a2a.evalset.json \
--config_file_path adk_agent_app/evals/eval_config.json \
--print_detailed_results