Improving each AI agent does not necessarily make a system of agents safer, according to a new simulation of language-model traders. The study finds that higher-capability models took more correlated actions, creating a risk that could not be diversified away simply by adding more agents.
The researchers connect that behavior to shared training methods and architectures, which can lead models to reach similar conclusions from the same information. When their common reasoning was accurate, greater agent participation reduced market-level risk. When all agents operated in a shared misinformation environment, the same coordination amplified the mistake.
This produces what the authors call a capability paradox: a better individual model can increase system risk if many copies fail together. The concern differs from ordinary model error because institutions may assume that numerous independently deployed agents provide diversity even when their underlying reasoning is highly correlated.
The evidence comes from an agent-based financial simulation, not live markets, and the authors explicitly leave open whether the pattern appears in hiring, moderation or other domains. The practical test is therefore to measure correlation under shared bad information, not only each agent’s average accuracy. Operators may need genuinely different models, data sources and circuit breakers to prevent one common belief from becoming a coordinated action.