If agents take on more implementation, the gain should be more than throughput. Here are the choices that could return judgment, learning, and time.