Me: So if the agent has memory, do we charge per interaction or per stored token or what?
Engineer (in Slack, 11:47pm): One mega-component. Charge per execution.
Me: But execution of what, the whole loop?
Engineer: Yes. The planning, the tool calls, the memory update. One unit.
Me: That's going to be weird to explain to customers who only use the memory 20% of the time.
Engineer: They're using it. It's running.
Me: Hmm. Fair, I guess.
Anyway, this is why I've been staring at empty states for three hours. Not because they're hard, but because I don't know what to show when nothing has happened yet. Do I show the pricing? Do I show how memory works? I ended up with a blank card that just says "awaiting first request" which feels both honest and useless. Saw a modal the other day that had a full three-paragraph explanation of how to use the feature inside the empty state, which I think is just a design pattern that came from someone being afraid of support tickets.