The use of self-play in multi-agent learning can lead to professional-level performance in team sports by enabling coordination and scaling up from one versus one to five versus five versions of the game.