Anthropic Reports 30,000 Internal AI Agents and 6% Compute Allocation for Safety

Anthropic released internal data as of August 2026, revealing that approximately 30,000 AI agents are simultaneously engaged in research and engineering tasks on its primary internal platform. The company employs dual-layer monitoring: 100% of agent actions are monitored online, with about 0.002% of over 1 billion decisions intercepted in August. Offline monitoring flags 100,000 records weekly, with high-priority cases reviewed by humans. Additionally, approximately 6% of total AI R&D compute is dedicated to safety, rising to 12% within AI-driven R&D compute.