AI agents aren't confidently wrong because of bad context — they're wrong because of bad data engineering