Anthropic has disclosed internal data on AI-driven automation in its R&D. As of August this year, Claude has led 26% of the company's AI R&D tasks, while this share was less than 1% in February, showing an astonishing growth rate.

Six-Level Automation Framework: Over 90% Achieve Human-Machine Collaboration, but No "Fully Autonomous"

Anthropic uses a six-level automation framework to assess the level of AI involvement. AL4 is defined as "AI-led": humans provide high-level goals, and AI completes most of the task process, with humans responsible for supervision and review. Currently, over 90% of R&D work reaches the level of human-machine collaboration or higher, but no R&D tasks have achieved full autonomy (AL5), and AI always remains under human supervision.

Internally, about 30,000 AI agents powered by Claude are working in parallel on research and engineering tasks. These agents can assign tasks to each other and correct errors crosswise, generating over 1 billion decisions in a single month in August.

80% of In-Repository Code is Generated by Claude, Participating in the Self-Improvement of the Next-Generation Model

To prevent risks, all AI operations undergo online safety checks before execution, and some abnormal decisions are directly intercepted. The background also regularly samples conversation records, and high-risk cases are submitted for manual review.

Code-level changes are more obvious: currently, over 80% of code merged into the repository is generated by Claude, and engineers' productivity has increased several times compared to the past. These AI agents take on a large amount of low-level R&D work, such as model experiments, code writing, data analysis, and result verification, and participate in the iteration of the next-generation Claude model, which is referred to in the industry as recursive self-improvement.