Developers often try to eliminate errors from AI agents. But science shows that being wrong — within reason — is what makes intelligence adaptive.