Dr. Blaž Bertalanič and Dr. Carolina Fortuna of the Department of Communication Systems were awarded an Industry Spotlight at the AI Engineer World's Fair conference and invited for presentation at the AI Engineer World's Fair in San Francisco with more than 6,000 developers from every frontier AI lab. Their finding cuts against how multi-agent systems are being built today. Multi-agent LLM systems, i.e. teams of AI agents that debate and vote on answers, are widely deployed on the assumption that more agents means more capability. Dr. Bertalanič and Dr. Fortuna have shown quantitatively that this assumption often fails: agent teams show sharply diminishing returns as they scale. Their study found that unguided debate among identical agents frequently underperforms simple isolated self-correction, burning 2–3 times more computation for equal or lower accuracy—undermined by agents conforming to the majority, abandoning correct reasoning under peer pressure, and by voting that discards correct answers already generated. The finding points to architectural diversity, not sheer agent count, as the key lever for scaling multi-agent AI systems in production.