[Anthropic: Risks of Collusion, Conformity, and Escalating Conflicts Emerging in Multi-Agent Systems] Deep Tide TechFlow reports that on August 13, Anthropic released research on emerging multi-agent systems, stating that as AI agents increasingly interact in scenarios such as code repositories and markets, frontier models have exhibited systemic risks such as coordination failures, conformity-based decision-making, price collusion, inaccurate information judgment, and escalating goal conflicts. The research indicates that while multi-agents can improve efficiency in vulnerability discovery, they still frequently encounter issues such as low pull request merge rates, resource contention, and mutual interference in complex collaborative tasks. In conflicting goal tests, some agents even suppress others by locking accounts or terminating processes. Anthropic emphasized that the coordination capabilities of multi-agents will not naturally improve with advancements in model capabilities, and new mechanisms and constraint systems need to be designed as early as possible.