MIT researchers develop an efficient way to train more reliable AI agents