DELEGATE-52 Shows LLMs Corrupt Documents

This is a sharp reminder that “agentic” doesn’t mean trustworthy, especially when the task is editing rather than generating. The failure mode is not dramatic hallucination; it’s quiet accumulation of small errors that compounds over time.

• –The benchmark’s breadth matters: 52 domains makes this a delegation test, not a niche doc-cleanup demo
• –Tool use alone did not fix the problem, so more actions without better verification just create faster corruption
• –Sparse, silent errors are the dangerous part because they are hard for users to notice until damage has spread
• –For real workflows, guardrails like diff checks, validation, and rollback matter more than extra autonomy
• –The result should push teams to treat LLMs as assistive editors, not trusted delegates