Hacker News discussion: Agentic Coding in the Wild: Characterizing GitHub Copilot Traces at Production

A new arXiv paper provides the first large-scale analysis of production workloads from AI coding agents like GitHub Copilot. Drawing on June 2026 traces spanning 3.2 million users, 13 million sessions, and 95 trillion tokens, the study finds that these tools create a distinct workload pattern—sparse user requests trigger long, autonomous chains of LLM calls coupled with tool use. This structure leads to high KV cache reuse within a single task but sharp drops across task boundaries, especially following model switches or context compaction. The paper also identifies long, variable idle times between user turns and demonstrates a lightweight predictor that can capture most of that idle time. The findings challenge current LLM-serving architectures and make an evidence-based case for infrastructure designed specifically for agentic workloads.