Monoculture Collapse¶
Monoculture collapse occurs when all agents in a multi-agent system share the same base model, causing systematic errors to propagate uniformly across the entire team (Reid et al., 2025). The team becomes a single point of failure at the calibration level regardless of how many agents it contains.
Mechanism¶
When all agents use the same foundation model, they share:
- The same training data biases
- The same knowledge gaps
- The same reasoning patterns
- The same calibration failures
If the base model systematically underweights a particular failure mode (because it was underrepresented in training data), every agent in the system underweights it. Adding more same-model agents provides no protection because they all share the same blind spots.
Nuclear Analogy¶
This maps directly to a well-understood nuclear safety concept: using identical equipment from a single manufacturer for all redundant safety trains. If a common design defect exists, all trains fail simultaneously. Physical defense-in-depth requires design diversity for exactly this reason. The same logic applies to AI reasoning layers.
Four Non-Resolutions¶
Several approaches appear to address monoculture but do not:
Stochastic variation: Different random seeds produce different outputs, but non-determinism is not independence. Varying outputs from the same probability distribution do not decorrelate systematic biases.
Chain-of-thought prompting: Requiring explicit reasoning does not change the underlying model weights. Different reasoning paths through the same biased model arrive at the same biased conclusions.
RAG with different sources: Providing different retrieved documents gives agents different knowledge but not different reasoning. The same model reasons about different facts with the same biases.
Fine-tuned variants: Models fine-tuned from the same base retain shared pre-training biases. Fine-tuning adjusts the surface but the deep structure remains correlated.
Implications¶
For systems where AI assessments are safety-credited, monoculture collapse means that multi-agent architectures using a single model family provide no additional reliability over a single agent. The apparent redundancy is illusory. True redundancy requires model-heterogeneity — different foundation models with genuinely different training and architectures, accepting the additional complexity this entails (see common-cause-failure).