Claude Now Leads 26% of Anthropic’s AI R&D Work, but Humans Remain in the Loop

Anthropic is increasingly using its own AI to build the next generation of its technology, with Claude now taking the lead on 26% of the company’s artificial intelligence research and development work. 

Invest in Indian Markets and Unlock Future Potential With 5paisa!

Open Account Now

That share has risen sharply from less than 1% in February. By August, more than 90% of Anthropic’s AI R&D involved work where Claude either collaborated with researchers or took the lead while remaining under human supervision. 

What it has not done is remove people from the process. 

Anthropic found no measured area of its R&D where Claude was working with complete autonomy. 

The company arrived at these figures through a prototype it calls the Anthropic R&D Automation Index. The system examines internal work records, including Slack activity and other documentation, and uses an automation scale developed by Epoch AI. 

That scale runs from AL0, where AI has no involvement, to AL5, where an AI system can work without a human in the loop. 

Claude’s 26% figure falls under AL4. At this stage, the AI can take a high-level instruction and complete most of the task, but a person still supervises the work. AL3 covers a more collaborative arrangement, where AI performs substantial parts of a task under closer human direction. 

Full autonomy would require considerably more. At AL5, an AI system would need to identify a problem itself, decide what work is required, build and test a solution and then deploy it without human involvement. Anthropic said none of the R&D activity it measured had reached that point. 

Thousands of AI agents are already working inside Anthropic 

The scale of AI use becomes clearer elsewhere in the company’s operations. 

Around 30,000 AI agents were carrying out research and engineering tasks at any given time on Anthropic’s most widely used internal platform in August. 

Those agents are monitored before their actions are executed. Anthropic said every action passes through an online monitoring system, usually within seconds. 

More than 1 billion agent decisions went through that process during August. About 0.002% were blocked, equivalent to roughly one in every 47,000 decisions. 

There is another layer of checks after the agents have acted. An offline monitoring system flags about 100,000 transcripts each week for further automated review. Of those, around 50 of the highest-priority cases are escalated to people. 

Anthropic said these systems are intended to pick up both immediate problems and patterns that may emerge gradually across multiple agents. 

Safety takes a larger share of compute in AI-led research 

The company has also disclosed how much computing capacity it is directing towards safety work. 

During a sample week between July 13 and July 20, about 6% of the computing power used across AI R&D went towards safety. Within R&D being driven by AI, that proportion was higher at around 12%. 

Anthropic plans to publish its new R&D automation metrics regularly, giving governments, researchers and the public a way to track how quickly AI is taking on more of the work involved in developing AI itself. 

For now, its own measurements show a substantial change in how that work is being done: Claude can lead a sizeable portion of research tasks, and tens of thousands of AI agents are already operating internally, but human oversight has not disappeared from the process.