Around 30,000 AI agents operate on Anthropic's main internal platform, producing more than 1 billion monitored decisions in August. The chatbot Claude completes most research and development tasks end to end from a high level prompt for 26% of AI R&D, while humans primarily supervise. More than 90% of measured R&D work involves AI human collaboration, though the company stated that none of the tasks are fully autonomous.

The share of tasks led by Claude rose from approximately 1% six months ago. Anthropic allocates 6% of its total AI R&D compute to safety, which increases to 12% for AI driven development. A real time safeguard blocked 0.002% of agent decisions in August, or about 1 in 47,000 decisions, to ensure oversight of the systems building future models.