Noise names the correlation problem itself: shared assumptions remove noise and leave the error; low disagreement among same-model agents is not evidence of accuracy.

The claim

Kahneman, Sibony and Sunstein grant that noise reduction can manufacture bias: "If all forecasters in a government office adopted the same unrealistically optimistic assumptions, their forecasts would not be noisy, but they would be wrong." Of algorithms they say: "Since they eliminate noise, they could be more reliably biased than human judges." Many agents on one base model are that government office: their agreement is low noise bought with shared assumptions. So the book does not count agent agreement as a virtue by itself; a collective has to measure accuracy on resolved questions, not agreement, and add judges whose errors differ. Inference beyond the quotes is Leo's.