AgentSysBench: Characterizing Agentic Workloads for AI Serving Systems

A recent study published on arXiv (2608.15127) presents AgentSysBench, a benchmarking suite and measurement toolkit aimed at analyzing agentic workloads—AI applications that involve long-duration tasks where LLMs manage tools, environments, and ongoing states, unlike traditional isolated model inference. The authors, whose names are not mentioned in the abstract, investigate ten typical agentic applications through a unified systems-level approach, integrating controlled deployments with production traces. The results identify six characteristics that set agentic workloads apart from standard LLM serving: execution is both heavyweight and stateful, with non-LLM elements causing latency in 5 out of 10 applications, and memory usage peaking at 28 GB per session. The research underscores that existing serving systems are based on outdated assumptions that do not meet the needs of these emerging workloads, offering essential insights for developing more effective serving infrastructure for agentic AI, particularly regarding latency, costs, and bottlenecks. The paper can be accessed on arXiv under the identifier 2608.15127, categorized as 'cross'.