Multi-agent systems: when splitting an agent helps and when it just costs more