Agents Break the Scheduler: Why ML-Based Resource Managers Fail for Agentic Workloads
Abstract
ML-based resource managers—learned autoscalers, utilization-driven schedulers, RL-based controllers—were designed for workloads with stationary resource profiles. AI agents executing tool calls via the Model Context Protocol (MCP) violate every assumption these systems rely on: resource demand is non-stationary within a single session, driven by model decisions rather than developer declarations, and invisible to the scheduler because tool-call telemetry flows to APM dashboards rather than infrastructure control planes. We characterize three concrete, measurable failure modes through experiments on a real MCP-based agentic system, and extend our evaluation with a controller shoot-out across six representative policies (including EWMA, a learned predictor, and Kubernetes HPA), multi-model validation across 10 LLMs from 9 families, cross-framework replication under LangChain, and cluster scale projection to N = 1,024 agents. The controller comparison shows that every existing approach sits on a bad violations-vs-waste frontier, while intent-driven renegotiation (AgentDRA) achieves 65× fewer violations than HPA and 36× less waste than static limits—because it sees the one signal current controllers lack: which tool the model has chosen next.