“Many historical and current approaches to multi-agent AI research are misguided because they are too heuristic and do not follow the "bitter lesson" of prioritizing scalable methods.”