I ran the same multi-agent experiment twice. Half the conclusions flipped.