Tom outlines that a multi-agent system should return with an interrupt step to handle asynchronous tasks.
