From Requests to Sessions: A Large-Scale Characterization of Human-Driven Agentic Workloads

Abstract

Coding agents are becoming an important source of LLM inference traffic, yet their workload behavior remains poorly understood. Existing workload studies largely focus on general LLM traffic, providing limited visibility into the characteristics of agentic workloads. To bridge this gap, we study 3.7 months of production traffic from FreeInference, spanning 209B input tokens and 12K sessions across 13 agent harnesses, revealing agentic workload characteristics that are not captured by request-centric views of LLM serving.

Our analysis shows that coding-agent sessions are long-lived and bursty, with tight LLM–tool loops separated by longer user idle periods. Contexts grow substantially while outputs remain short, causing cache reads to dominate token volume and API cost. Tool execution dominates active-session latency, yet tool parallelism remains uncommon. Prefix reuse is high but disrupted by user idle gaps and context mutations performed by agent harnesses. A 15-minute KV-cache retention time captures most practical reuse, while longer retention provides diminishing returns at substantially higher storage cost. We will release the trace alongside the paper to enable further research on agentic workloads and serving systems.