The primary bottleneck in agentic AI is no longer capability. It's entropy. When agents "do what you didn't ask," it is rarely the result of a bad prompt. It's that agents fill gaps with assumptions, leading to silent failures that users eventually stop checking. This session breaks down the design discipline required to move from experimental demos to production-grade reliability. We will examine why "no design" is worse than "bad design," and why human-in-the-loop only works when humans run tasks independently rather than merely validating agent output.