Microsoft researchers released a large-scale telemetry study showing GitHub Copilot now initiates most of its own processing, with roughly 87 percent of large language model calls occurring after a single developer prompt. Analyzing 13.5 million sessions from Visual Studio and Visual Studio Code, they observed agent workflows where model and tool calls occur in near-equal ratios, often forming extended loops of reasoning, file access, builds and terminal commands. Deep retry loops after failures can quadruple compute demands, while most activity centers on reading and exploration rather than direct editing, increasing infrastructure planning questions without yet addressing Copilot’s task success rates or coding quality.
This update represents a notable development in the Ai sector. Organizations and founders tracking this space should evaluate potential strategic and technical implications on their operations.