Scott Abel, writing in The Content Wrangler, makes a point that should humble anyone building AI agents. Claude Shannon described the core problem with generative AI back in 1950. His paper "Programming a Computer for Playing Chess" wasn't really about chess. It was about what happens when a machine faces too many possibilities and not enough compute to evaluate them all. The machine has to guess. Sound familiar? Shannon didn't demand perfection from his chess program. He wanted "tolerably good" performance. That pragmatic stance holds up. Modern LLMs work the same way. They don't know answers. They predict what a good answer looks like, token by token. When signals are strong, the output works fine. When signals are weak or missing, the model still produces something polished and confident. It just might be wrong. Abel highlights a concept called "processing fluency," where people judge fluent, easy-to-read language as more likely true. AI systems exploit this constantly. The response flows nicely. It uses the right jargon. Reads like it belongs in your documentation. But coherence isn't accuracy. A response can sound authoritative while quietly steering users into a ditch. Psychologists have found that statements simply easier to read get judged as more true. Dangerous bias when your AI produces smooth prose regardless of factual grounding. For anyone deploying AI agents, Shannon's framework still matters. Signal quality determines output quality. Without proper structure and clear boundaries, agents fill gaps with confident nonsense. They don't refuse to answer. They infer and smooth over missing information, like someone nodding along in a conversation they stopped understanding eight minutes ago. Better models won't save you. Better inputs will.