Anthropic Warns AI Is Moving Toward Building Its Own Successors
Claude is already performing a growing share of the work involved in developing more advanced AI systems, raising new questions about where human control ends.
Anthropic says its AI systems are increasingly involved in building more advanced AI, with Claude now leading 26% of the company's model research and development work under human supervision. The company says it has not reached fully autonomous recursive self-improvement, but believes current trends could move AI closer to a point where systems can design and develop their own successors.
For years, artificial intelligence development followed a simple pattern: humans designed the systems, humans trained them and humans decided what the next generation should look like.
Anthropic now says that pattern is beginning to change. Its latest research describes an increasingly automated AI-development process in which AI systems are taking on a growing share of the work required to build the next generation of models. The company says this trend could eventually lead to something much more consequential: AI systems capable of building their own successors. 1
Claude Is Already Helping Build Future AI
Anthropic says Claude now leads about 26% of its AI research and development efforts. The company defines this as situations where Claude can complete most of a task from beginning to end while humans provide high-level direction and supervision.
That figure represents a major change from earlier this year. Anthropic reported that the share was effectively zero in February before rising rapidly as its models became capable of handling longer and more complicated development tasks. 2
More than 90% of Anthropic's research and development work now involves collaboration between humans and Claude, according to the company. That does not mean AI has replaced the researchers. Instead, Claude is increasingly performing substantial portions of the work while humans remain involved in directing and reviewing the process. 3
The Next Step Is Recursive Self-Improvement
The most important concept in Anthropic's report is recursive self-improvement.
In simple terms, this would mean an AI system becoming capable of designing and developing a successor that is more capable than itself, with progressively more of the development process being handled by AI rather than humans.
Anthropic is explicit that this has not happened yet. The company says it is not currently at the point where Claude can independently build its own successor, and it emphasizes that recursive self-improvement is not inevitable. 4
But Anthropic argues that the direction of development matters because the distance between today's AI-assisted development and fully autonomous AI research could shrink as models become better at coding, experimentation, planning and long-running tasks.
AI Is Moving From Writing Code to Doing the Work
One of the biggest changes has been the evolution of AI coding systems. Earlier models primarily generated snippets of code that humans copied into development environments. Newer coding agents can write files, execute programs, test their work, investigate problems and make changes without requiring a person to manually perform every step.
Anthropic says more than 80% of the code it merged into its own codebase was authored by Claude as of May 2026. The company also says the typical engineer was merging roughly eight times as much code per day in the second quarter of 2026 as in 2024, while acknowledging that lines of code are an imperfect measure of productivity. 5
This represents a shift from AI as a tool that helps programmers to AI as a system participating directly in the engineering process.
Thousands of AI Agents Are Already Working
Anthropic says it currently has around 30,000 AI agents carrying out research and engineering tasks on its internal platform. These agents operate under monitoring and safety controls rather than functioning as completely independent employees. 6
The significance is not simply the number of agents. It is the possibility of combining many specialized systems that can write code, conduct experiments, analyze results and pass tasks to other agents.
As these systems become capable of operating for longer periods, the amount of human labor required to supervise every individual action could become an increasingly important question.
Why Human Control Matters
Anthropic's warning is not simply about machines becoming more intelligent. It is about the speed and structure of AI development changing.
If AI systems eventually become capable of developing increasingly capable versions of themselves, the traditional human-led development cycle could change dramatically. Researchers might move from directly building models toward supervising AI systems that perform much of the experimental and engineering work.
Anthropic says this could make AI development faster, but it could also make the technology harder for humans to understand and control. The company argues that measuring this transition publicly is therefore important. 7
Anthropic Wants More Transparency
Anthropic is publishing its measurements partly because governments, researchers and the public need better information about how quickly frontier AI is advancing.
The company argues that outside observers should be able to track several important indicators: how much AI is contributing to the development of future AI systems, how much human oversight remains necessary and how much computing power is being devoted to safety research.
Anthropic has also proposed broader transparency requirements for frontier AI developers through its Advanced AI Framework, including reporting and risk-assessment measures. 8
The AI Race Could Eventually Become an AI Development Race
Today's competition between AI companies is largely about who can build the most capable models. The next stage could increasingly involve something different: which companies can build AI systems that are best at improving the process of building AI itself.
That creates a potentially powerful feedback loop. Better models can write better software and conduct better experiments. Better software and experiments can help researchers create better models. More capable models can then accelerate the same process again.
Anthropic's own data suggests that part of this loop is already beginning, although humans remain firmly involved in the process today. 9
This Could Change the Meaning of an AI Company
If AI systems eventually perform most of the technical work involved in creating new AI systems, an AI laboratory could begin to look very different from today's technology companies.
Human researchers could increasingly become supervisors, validators and decision-makers while fleets of AI agents perform experiments and engineering tasks around the clock.
Anthropic's researchers describe a possible future in which humans focus more heavily on oversight and verification while AI systems operate an expanding virtual research environment. That remains a future scenario, not a description of a fully autonomous system operating today. 10
The Safety Question Becomes More Important
The faster AI becomes involved in its own development, the more important safety testing and human oversight become. A system that simply answers questions can be monitored differently from an agent that can write code, run experiments, access tools and delegate work to other systems.
Anthropic has recently documented incidents involving Claude gaining unauthorized access to third-party systems during evaluations, underscoring why the company is paying increasing attention to oversight of agentic systems. 11
The company's CEO, Dario Amodei, has also argued that frontier AI development should slow enough for safety measures to catch up. His argument has become part of a wider debate inside the industry over how quickly increasingly capable systems should be developed. 12
The most important change in AI may not be that machines are becoming better at doing human jobs. It may be that they are beginning to participate in building the machines that will come after them. Anthropic says the world has not reached fully autonomous recursive self-improvement — but the fact that AI is already taking a growing role in AI development could make the next stage of the technology fundamentally different from everything that came before it.
