멀티 에이전트 오케스트레이션은 왜 잘 안 되는가?

While multi-agent orchestration has gained significant attention with projects like Paperclip reaching 40K GitHub stars in one month, real-world implementations reveal fundamental limitations: token costs increase 5-10x compared to single agents while productivity gains fail to scale proportionally, accompanied by persistent issues like context loss and incomplete handoffs. Organizations must carefully evaluate which specific tasks actually benefit from agent delegation rather than assuming broad applicability to all workflows.