Differences in Alignment Behaviour between Single-Agent and Multi-Agent AI Systems — LessWrong