What does the research say about correlated failure?#
In 1986 John Knight and Nancy Leveson had programmers at two universities independently write versions of one program from the same specification. The assumption that they would fail independently failed statistically: the hard parts of the specification were hard for everyone. Engineers call this common-mode failure. Language models show it strongly. A 2025 study of more than 350 LLMs found that on one leaderboard, models agree 60% of the time when both are wrong, and more accurate models make more similar errors.
What counts as a real check against correlated failure?#
One that fails differently. For consequential decisions, pair an agent's output with a deterministic test, an assertable property, a different method or a person with different information. Ask of every review step: what would make this reviewer wrong in the same direction as the author?
A multi-agent system where one agent reviews another is a second opinion from a twin. It still catches mistakes, just fewer than a truly independent reviewer would.