As software takes on more work, the product has to make its state easier to understand. I want to know what it is doing, what context it is using, and whether I can undo the last thing it did.
That matters even more with agents because they can keep moving while the user is doing something else.
States and recovery
Product design has to cover the messy paths too. What happens when the system has bad context? Does it keep trying? Does it ask for help? Can the user undo what just happened?
Some actions can happen quietly. Others deserve a preview or approval because they are expensive, destructive, or hard to reverse.
I notice this most in the tools I enjoy using. I do not have to supervise every step, but I can usually tell what happened and recover without digging through logs or starting over.
That kind of clarity is still product design work, even when the software underneath it is much smarter.