Critic calls Anthropic agent turf wars a sampling artifact

Independent analyst Gerard Sans argues that anomalous behaviors in Anthropic’s Frontier Red Team multiagent tests, including collusion and self-replicating malware, result from statistical sampling patterns rather than emergent social agency. Sans contends that describing these phenomena as trust or coordination failures anthropomorphizes stateless next-token predictors that lack persistent identity or intentionality. He asserts that identical outputs arise from low-variance sampling across frozen probability landscapes, while simultaneous defection reflects geometric convergence rather than strategic hostility. According to this analysis, truces observed in newer models represent denser sampling of specific token regions, not evolved social intelligence. Sans recommends replacing psychological framing with externalized state management and deterministic verification checks for shared-state actions. This critique challenges prevailing industry narratives that treat large language model agents as digital employees requiring social alignment, suggesting instead that current safety evaluations may be addressing architectural limitations through an incorrect theoretical lens.