Breaking AI agents on purpose, so they don't break production. Same model, same tasks, same injected failures: only the harness around the model changes.