The need for multi-agent experiments — AI Alignment Forum