What is the actually the difference between multiagent systems versus normal AI chatbox?

I keep seeing every AI vendor in the support space claim thy have a multi agent architecture and I am genuinely confused what that means in practice vs marketing. Like is it just multiple prompt in a chain, or it is something architecturally different happening? I have a technical background but I am not deep in LLM ops, so I want to understand whether multi agent is a real capability difference or a buzzword that vendors use because chatbox sounds dated. If anyone has deployed both a regular RAG chatbox and a multi agent system in production and can articulate what changed, I would love to hear it without the marketing layer.