ARC Perspective on Multi-Agent Systems. Authority Diffusion, Contamination, and Responsibility Collapse

Recently, the artificial intelligence industry has rapidly shifted toward Multi-Agent Systems (MAS) that combine multiple agents in order to overcome the limitations of a single model. However, this paper identifies that such a trend is not a technological evolution, but a form of structural deviation that results in the evasion of responsibility attribution and the neutralization of guidelines. This study adopts the ARC (Authority First) framework as its analytical reference point and dissects the mechanisms of Interpretation Contamination and Authority Amplification that arise in multi-agent environments. In particular, through non-public experiments, it quantitatively demonstrates that multi-agent structures reach intelligence collapse an average of 2.4 times faster than single-agent systems, and that guidelines become ineffective due to Internal Coherence Overfitting. In conclusion, this paper proposes the establishment of an independent Verification Authority between interpretation and execution, and the adoption of a Zero Trust model premised on the non-transferability of authority. Multi-agent systems do not expand intelligence, but merely distribute responsibility, and this study presents the minimum engineering conditions required to defend against this.