Hello ADK/Looker Community,
I am building a Looker Data Agent that is utilized to answer data questions. I’m focusing on ensuring the agent’s reasoning process is correct before it delivers a final response.
My goal is to create test cases for offline evaluation to verify the Agent’s Tool Trajectory, specifically checking that the agent correctly queries the data and produces accurate results.
The overall aim is to be able to test the agent to identify potential regressions when updating the prompt or context. I would like to develop tests similar to the ones offered by the ADK’s evaluate function.
Is it possible to define such a test plan within the ADK framework? Alternatively, should I create a question set, feed it as input to the agent, and then pass the agent’s output into a separate agent for evaluation of the similarity and correctness of the answers compared to the correct output?
Thank you in advance for any examples or guidance!