Orchestrating multiple AI agents on a puppeteer workflow—where does coordination actually break?

I’ve been reading about autonomous AI teams and the idea of having multiple agents work together on a complex task. Like, one agent handles navigation, another handles data extraction, another aggregates results. It sounds powerful in theory, but I’m wondering where the practical limitations are.

What happens when agent A needs something agent B hasn’t finished yet? Or when the navigation changes mid-task and the extraction agent doesn’t know how to adapt? How do you actually coordinate them without everything collapsing into chaos?

Has anyone built something like this for a real workflow? Where did you run into trouble?