They are arguably "fun", it's kind of like functional programming but in NLP. We found multi-agent systems to work well for workflows where there's lots of unstructured inputs. E.g. a "content writer" that takes resolved support tickets and turns them into an update against documentation if there's something novel to update. We tried that with a structured workflow and it didn't work very well/reliably, while a multi-agent approach worked well. For more traditional mostly deterministic workflows, I agree, there's no need.