Position: Safety and Fairness in Agentic AI Depend on Interaction Topology, Not on Model Scale or Alignment

System structure matters more than individual model quality: The paper challenges the common assumption that using better or more carefully aligned AI models automatically makes multi-agent systems safer. Instead, how agents interact with each other—the "interaction topology"—is the dominant factor determining whether the system behaves safely and fairly.

Three specific failure modes emerge from how agents are connected: The research identifies ordering instability (results change based on which agent goes first), information cascades (early decisions bias later ones regardless of accuracy), and functional collapse (systems appear fair on metrics while losing the ability to distinguish between risky and safe decisions). These problems persist across different model types and sizes.

Scaling up models can actually worsen these problems: Counterintuitively, using more capable AI models strengthens these topology-driven failures by making agents more likely to reach consensus and by making initial decisions harder to correct—the opposite of what intuition might suggest about safety improving with better models.

Current safety evaluation methods are inadequate: Standard approaches to testing and aligning individual AI models cannot detect these interaction-based failures, meaning systems could pass safety reviews while still exhibiting dangerous behavior when deployed as multi-agent systems in high-stakes decisions.

Regulation should focus on system architecture, not just model capability: Before deploying agentic AI systems in important decisions, organizations should be required to test how their systems behave under different interaction structures and demonstrate robustness across various architectural designs.