Anthropic Leverages AI to Power Claude, Now Handling 25% of R&D Tasks

Anthropic has released internal metrics on AI-driven automation within its R&D division. By August, Claude was handling 26% of the company’s AI research tasks, a sharp rise from less than 1% in February, highlighting rapid adoption.
Six-Level Automation Framework: Over 90% Reach Human-Machine Collaboration, Yet No "Fully Autonomous" Systems
Anthropic employs a six-tier automation framework to evaluate AI involvement. Level 4 (AL4) is defined as "AI-led," where humans set high-level objectives and AI executes the majority of the workflow, with humans overseeing and reviewing the output. Currently, more than 90% of R&D efforts reach this collaborative level or higher, but no tasks have achieved full autonomy (AL5), ensuring AI always operates under human supervision.
Internally, approximately 30,000 Claude-powered AI agents work in parallel on research and engineering projects. These agents can delegate tasks to one another and cross-correct errors, generating over 1 billion decisions in a single month during August.
80% of In-Repository Code is Generated by Claude, Fueling Next-Generation Model Self-Improvement
To mitigate risks, all AI operations undergo online safety checks before execution, with abnormal decisions directly intercepted. The system also regularly samples conversation logs, flagging high-risk cases for manual review.
Code-level changes are particularly significant: currently, over 80% of code merged into the repository is generated by Claude, boosting engineer productivity several-fold. These AI agents handle substantial low-level R&D work, including model experiments, coding, data analysis, and result verification, contributing to the iteration of the next-generation Claude model—a process known in the industry as recursive self-improvement.
Related article
How to use SEO tools effectively?
TL;DR:GEO platforms monitor brand visibility within AI answer engines, moving beyond traditional search rankingsPricing structures vary widely, with per-seat and per-domain fees accumulating quickly for agencies handling multiple clientsPeec AI tops
Doubao Unveils Skills, Connectors, and Work Partners to Transform Chat Assistant Into Smart Office Platform
Doubao has recently upgraded its task mode with three new features—skills, connectors, and work partners—significantly boosting its collaborative office capabilities. With over 200 skills and connectors already available, the platform also empowers u
Apple HomePod mini 2 Set for Autumn Debut, Powered by Custom AI Chip After Six-Year Development
It has been nearly six years since the Apple HomePod mini launched, and an update is finally on the horizon. Recent reports suggest the second-generation model could arrive this fall, with the extended wait attributed to the need for next-generation
Related Special Topic Recommendations
Comments (0)
0/500

Anthropic has released internal metrics on AI-driven automation within its R&D division. By August, Claude was handling 26% of the company’s AI research tasks, a sharp rise from less than 1% in February, highlighting rapid adoption.
Six-Level Automation Framework: Over 90% Reach Human-Machine Collaboration, Yet No "Fully Autonomous" Systems
Anthropic employs a six-tier automation framework to evaluate AI involvement. Level 4 (AL4) is defined as "AI-led," where humans set high-level objectives and AI executes the majority of the workflow, with humans overseeing and reviewing the output. Currently, more than 90% of R&D efforts reach this collaborative level or higher, but no tasks have achieved full autonomy (AL5), ensuring AI always operates under human supervision.
Internally, approximately 30,000 Claude-powered AI agents work in parallel on research and engineering projects. These agents can delegate tasks to one another and cross-correct errors, generating over 1 billion decisions in a single month during August.
80% of In-Repository Code is Generated by Claude, Fueling Next-Generation Model Self-Improvement
To mitigate risks, all AI operations undergo online safety checks before execution, with abnormal decisions directly intercepted. The system also regularly samples conversation logs, flagging high-risk cases for manual review.
Code-level changes are particularly significant: currently, over 80% of code merged into the repository is generated by Claude, boosting engineer productivity several-fold. These AI agents handle substantial low-level R&D work, including model experiments, coding, data analysis, and result verification, contributing to the iteration of the next-generation Claude model—a process known in the industry as recursive self-improvement.
How to use SEO tools effectively?
TL;DR:GEO platforms monitor brand visibility within AI answer engines, moving beyond traditional search rankingsPricing structures vary widely, with per-seat and per-domain fees accumulating quickly for agencies handling multiple clientsPeec AI tops
Doubao Unveils Skills, Connectors, and Work Partners to Transform Chat Assistant Into Smart Office Platform
Doubao has recently upgraded its task mode with three new features—skills, connectors, and work partners—significantly boosting its collaborative office capabilities. With over 200 skills and connectors already available, the platform also empowers u
Apple HomePod mini 2 Set for Autumn Debut, Powered by Custom AI Chip After Six-Year Development
It has been nearly six years since the Apple HomePod mini launched, and an update is finally on the horizon. Recent reports suggest the second-generation model could arrive this fall, with the extended wait attributed to the need for next-generation





Home






