What happens when AI systems start depending on each other?

One AI uses another AI's output. That output gets passed to another system, which makes a decision and triggers an action.

At what point does the whole chain become too difficult to understand or control?

Is the real challenge of AI no longer the model, but everything connected to it?