setting agents loose on some refactoring experiments gives a quick feel for whether architecture changes will shake out cleanly or turn ugly