A critic agent is built using the same model family as the generator agent it reviews. Which failure mode does this configuration make more likely?
Answer: Both agents will share the same systematic biases, causing the critic to miss errors the generator consistently makes
When critic and generator share the same model family, they tend to share the same training-induced biases and blind spots. A systematic error that the generator makes consistently — for example, a category of factual mistakes or a style of overconfident reasoning — will not be caught by a critic trained on the same distribution. Using heterogeneous model families for critic roles is the mitigation: different training distributions reduce the probability that both models share the same blind spot. The other options describe incorrect or irrelevant concerns — model families do not communicate internally, and token costs are unrelated to model family similarity.