Describe your agent and get a test suite for it: 4 to 6 quality dimensions with pass thresholds, 15 to 25 test cases written as real inputs, a grader for each case, rubrics for anything that needs an LLM judge, and a Python script that runs the suite and prints a scorecard. The script exits with an error when a threshold fails, so you can run it in CI after every prompt or model change.