Pulkit Yadav removes a cancellation tool from an airline-support assistant. All three previously passing tasks that require it fail. Yet the overall score rises from 25 to 27 successes across 50 tasks, because unrelated outcomes change.
The pilot inserts seven controlled defects into an agent, a model running multi-step tasks with tools. It compares hand-written and generated tests, then repeats tool removal and forced human escalation in retail support.
- Under a rule requiring a decline greater than ten percentage points and a statistical check, generated airline tests detect one of seven defects. Hand-written tests detect none.
- A defect affecting only a few baseline-passing tests cannot directly move the overall score very far. The paper derives that ceiling before evaluating the defective configuration.
- Scoring affected requirements separately exposes the missing tool. Tracking behavior is also necessary: forced escalation can escape requirement labels and final-state checks.
- In retail, removing a tool breaks all six previously passing dependent tasks. Total successes nevertheless rise from 20 to 21 out of 48.
The initial airline example is less stable than it first appears. Diagnostic reruns give the original assistant only three successes across 21 affected-task trials, all on one task. The tool-removed version passes none. The tool is broken, but the initial baseline passes were noisy.
Caveat: This is a mostly single-run pilot using the same model as assistant, simulated user and grader. Scoped scoring and diversity analyses are post-hoc, and the ten-point threshold changes the conclusions.
Keep the relationship between each requirement, its tests and the behavior observed. A green average can hide a business operation that no longer works. Proposed learned behavior-grouping methods remain untested; the paper’s demonstrated contribution is narrower.