Why do single-stream, single-turn benchmark evaluations substantially overestimate baseline serving performance in agentic deployments?

In traditional Mixture-of-Experts (MoE) serving engines, generation throughput drops by more than what percentage when transitioning from single-turn reasoning to multi-turn coding tasks?

Analyze why static offloading baselines deteriorate in multi-turn agentic serving environments, and explain how advanced expert management techniques achieve stable decode throughput across extended sessions.