Inside OpenAI: How AI Coding Agents Accelerate Internal Research

Simon Willison analyzed internal OpenAI data revealing a massive surge in the adoption of AI coding agents by researchers throughout 2026. Daily expenditure on AI agents per median researcher climbed sharply from near zero early in the year to approximately $600 by late August 2026. This trend illustrates how frontier AI labs are operationalizing agentic engineering to accelerate their own workflows, bringing theoretical recursive self-improvement into practical application. As AI models help design and build their successors, the feedback loop for AI research and development is rapidly shortening. The usage chart shows a particularly steep increase in compute spend starting in late July 2026, which Willison speculates coincided with internal employee access to unreleased models like GPT-6 Astra. The report aligns with recent writings from OpenAI leadership, including Chief Scientist Jakub Pachocki, explicitly framing internal workflows around recursive self-improvement (RSI).