[Anthropic: Approximately 30,000 AI agents working simultaneously, 6% of AI R&D compute allocated to safety] Anthropic has released internal data measuring the pace of cutting-edge AI development, stating that as of August 2026, its most commonly used internal platform had approximately 30,000 AI agents engaged in research and engineering tasks at any given time. The platform implements dual-layer supervision for these agents, both online and offline: 100% of agent operations are monitored online before execution, and among over 1 billion decisions analyzed in August, approximately 0.002% were intercepted, equivalent to about 1 in every 47,000. Offline monitoring flags around 100,000 records weekly, with about 50 high-priority records identified by classifiers and sent for manual review. In terms of compute allocation, Anthropic's sampling statistics indicate that approximately 6% of AI R&D compute is dedicated to safety efforts; within AI-driven AI R&D compute, the proportion allocated to safety work is around 12%.