What does the research say?#
Knight and Leveson showed in 1986 that even independently written programs fail together more than chance predicts. Their experiment used separate teams, separate implementations and a common specification, and the failures still clustered. Agents sharing a base model, a prompt and a context have far more in common than that.
What counts as a real check?#
For decisions that matter, the check should work in a different way: a deterministic test, a different method, or a person with different information. The question to ask of any review step is what it would take for the reviewer to be wrong in the same direction as the thing being reviewed.
This is the structural argument for keeping a human in some paths, and it is also why a multi-agent system with a reviewer role should not be counted as two independent opinions.