Early agent systems will make mistakes. The goal should be learning responsibly, not pretending the mistakes do not matter.