Some of the most prominent AI companies say they are approaching “recursive self-improvement,” a long-theorized capability where AI autonomy builds and designs their own successors, a development that AI executives say could accelerate scientific progress but also has ignited recent calls to slow AI’s advance.
From Anthropic’s Claude to OpenAI’s automated research intern, the race is moving toward models that write code, run tests, review results, and support the next generation of AI systems.
AI Autonomy Labs Move Toward Self-Improving Models
Anthropic said its Claude models are now leading 26% of the company’s model research and development work, meaning Claude can complete most of a given task “end-to-end from a high-level prompt,” while still operating under human supervision.
The Claude-parent said its models have not yet reached a full AI autonomy level but encouraged competitors to disclose similar metrics.
Once treated as a distant risk, recursive self-improvement is now becoming a live question inside the world’s top AI labs. The concept refers to AI systems that help improve their own design, then use those gains to develop even stronger successors.
It’s a new phase for AI autonomy, where models are no longer limited to answering questions or writing simple code. Instead, they are becoming active participants in research, testing, and engineering.
“The really important thing here is that as AI is doing more of it, it gets faster, because AI operates just much, much more quickly than the humans do,” said Anthony Aguirre, president and CEO of the nonprofit, Future of Life Institute, warning that speed is the central risk.
According to Aguirre, autonomous recursive self-improvement would mean AI systems designing successive version of themselves at a speed no longer limited by human speed, calling the pursuit “probably the worst idea in the history of humanity” even as companies press ahead with it.
The AI agent autonomy levels spectrum is now a governance issue, as each new level gives systems more room to plan, execute, and correct tasks with less human input, highlighting that the risk is not only technical.
OpenAI has also said it developed an automated “research intern” that can carry out well-defined research tasks under human direction. The company said it is working toward an automated AI “researcher” by March 2028.
That target shows how quickly agentic AI autonomy is moving from lab theory to practical deployment. It also explains why companies are trying to measure how much of their own research is now being handled by models.
Control Becomes the Real Test
The rise of levels of AI autonomy is dividing the industry between those who see self-improving systems as a path to scientific progress and those who fear a loss of human control.
Anthropic said there is still no point where Claude works “fully autonomously” without humans involved.
This is where the AI autonomy core of the debate sits: whether humans can safely benefit from faster model development without handing too much control to machines built to improve themselves.
“We have already, for years, been using these models in supportive roles for creating the next version of these models. So people use the past generation of models to write code for the AI systems that then create the next generation,” said John Thickstun, an assistant professor of computer science at Cornell University, highlighting that AI has already been used in supportive roles for years.
Still, the move from support to leadership is what worries researchers. If AI autonomy keeps increasing, monitoring the system may become harder just as its decisions become more important.
The AI agent autonomy levels spectrum also matters for regulators, because not all self-improving systems carry the same risk. A model that suggests code changes is different from one that designs experiments, launches agents, and evaluates its own success.
For developers, agentic AI autonomy could cut the time needed to test new architectures or find flaws. For the public, it raises a harder question: who is responsible when machines help build machines?
The emergence of autonomous software engineering agents gives AI labs more speed but also gives governments less time to understand what is happening inside private research systems.
That is why the levels of AI autonomy debate is moving beyond Silicon Valley. It is now tied to safety, transparency, and whether frontier developers should publish regular metrics on how much of their work is being done by AI.
Anthropic said other developers could share similar measurements using public methods. It argued that such reporting would give governments and third parties better visibility at the pace of frontier AI development.
The AI autonomy core challenge is simple but unresolved: the more useful these systems become, the harder they may be to slow down.
For now, AI autonomy remains under human supervision. But with self-directed AI solutions already entering research workflows, the question is no longer whether AI can help build better AI, but how much control humans are willing to give away.
Inside Telecom provides you with an extensive list of content covering all aspects of the tech industry. Keep an eye on our Intelligent Tech sections to stay informed and up-to-date with our daily articles.
