MAS-H2: A Hierarchical Multi-Agent System for Holistic Cloud-Native Autoscaling

Researchers Hamed Hamzeh and Parisa Vahdatian have introduced MAS-H2, a novel hierarchical multi-agent system designed to solve the 'strategic void' in cloud-native platforms like Kubernetes. Current autoscaling tools like the Horizontal Pod Autoscaler (HPA) and Cluster Autoscaler (CA) are reactive and metric-driven, often leading to resource waste and performance issues. MAS-H2 addresses this by decomposing the control problem into three intelligent layers: a Strategic Agent that encodes business goals (like cost vs. performance), Planning Agents that create joint, proactive scaling plans using forecasting, and Execution Agents that carry out the plans.

The team built and tested a MAS-H2 prototype as a Kubernetes Operator on Google Kubernetes Engine (GKE). Under a predictable 'Heartbeat' workload, MAS-H2 kept application CPU usage under 40%, resulting in over 50% less sustained CPU stress than the native HPA baseline, which typically operated above 80%. In a more volatile 'Chaotic Flash Sale' scenario, the system's proactive planning filtered transient noise, deployed more replicas, and reduced peak CPU load by 55% without causing under-provisioning. Beyond raw performance, MAS-H2 also demonstrated the ability to perform a zero-downtime strategic migration between cost- and performance-optimized infrastructures, showcasing its holistic management capabilities.