Microsoft and the University of Illinois Urbana-Champaign analyzed 13.5 million GitHub Copilot sessions from 3.2 million users, representing the first large-scale empirical study of AI coding agents in production. During one week of data collection in June 2026, Copilot generated 761 million LLM calls, 775 million tool invocations, and processed 95 trillion tokens. The study reveals that 87% of LLM calls are agent-initiated rather than user-initiated, and KV-cache hit rates drop significantly when users issue new turns or switch models. Researchers identified five distinct user archetypes with a 50-fold difference in token consumption between the lightest and heaviest users, and discovered that retry loops can amplify compute demands by up to four times. These findings advocate three infrastructure shifts: moving from per-request to per-turn or per-session scheduling, accounting for retry amplification, and making idle-time prediction a first-class infrastructure concern.
Source: Read the original article
