Learning by building a multi-agent design lab. Yet, single agent works better