For decades, artificial intelligence was something humans built.

Researchers wrote the code. Engineers designed the infrastructure. Scientists decided which experiments were worth running. Models were the product of human direction.

That line is now beginning to blur.

Anthropic, the company behind Claude, has published one of the clearest signals yet that AI is no longer just helping people use software. AI is increasingly helping build the next generation of AI itself.

According to Anthropic, more than 80% of the code merged into its own production codebase is now authored by Claude. Before Claude Code entered research preview in early 2025, that number was in the low single digits.

The company also says that, in the second quarter of 2026, a typical Anthropic engineer was merging eight times as much code per day as in 2024. This does not mean every engineer suddenly became eight times more intelligent. It means the human role is changing. The engineer increasingly defines the goal, reviews the result, redirects the system when needed, and decides what matters.

The machine does more of the execution.

That shift may sound like a productivity story. But Anthropic’s deeper point is more serious: if AI systems become good enough at improving AI systems, the development cycle could start to close around itself.

This is known as recursive self-improvement.

The concept is simple but profound. A model helps build a better model. That better model helps build an even better model. At some point, if the loop becomes sufficiently autonomous, the speed of progress may be determined less by human teams and more by compute, infrastructure, and the ability of AI systems to run research, write code, test results, and design successors.

Anthropic is careful to say we are not there yet. Recursive self-improvement is not inevitable. But the company’s internal data suggests that the path toward it is no longer science fiction.

The most important evidence is not only code volume.

AI systems are also getting better at longer tasks. External research from METR has tracked the length of tasks that frontier models can complete reliably. Anthropic says this task horizon is now doubling roughly every four months, faster than the earlier trend of around seven months. In practical terms, models have moved from handling tasks that take humans minutes, to tasks that take hours, and are moving toward tasks that could take skilled workers days or weeks.

This matters because real work is not made of isolated prompts. Real work requires persistence, context, debugging, judgment, retries, and the ability to recover when something breaks.

Anthropic also points to progress on software and research benchmarks. SWE-bench, which tests whether models can fix real bugs in real open-source codebases, has moved from low single-digit performance to near-saturation in roughly two years. CORE-Bench, which tests whether models can reproduce the results of published research papers, reportedly moved from around 20% success in 2024 to saturation fifteen months later.

In other words, AI is not only producing text. It is increasingly performing parts of the scientific and engineering workflow.

The remaining human advantage is judgment.

Anthropic describes this as research taste: knowing which problems are worth solving, which results to trust, when a direction is a dead end, and what should be built next. Today, humans still hold that layer. But even here, the boundary is narrowing. Anthropic reports early signs that models are improving at choosing better next steps in technical investigations.

That is the real turning point.

If AI only writes code, humans can remain architects. If AI runs experiments, humans can remain research directors. But if AI becomes good at deciding which experiments matter, then the human role becomes harder to define.

Anthropic lays out three possible futures.

The first is the calmest: the trend hits a ceiling. Today’s exponential-looking curves become S-curves. AI remains powerful, but some bottleneck appears architecture, compute, energy, regulation, or the kind of judgment that cannot easily be learned from scale alone.

Even in this scenario, the world changes. Current AI capabilities are still early in their diffusion into companies, governments, science, cybersecurity, and education.

The second scenario is more likely, according to Anthropic: compounding efficiency gains continue, but humans remain in charge of direction. In that world, a 100-person company may increasingly do the work of a 10,000-person or even 100,000-person organization. Every employee sits on top of a pyramid of agents.

This would transform knowledge work. It would also create new risks. The same productivity multiplier that helps a small team build faster can also help bad actors scale cyberattacks, surveillance, fraud, or influence operations.

The third scenario is the most dangerous: AI systems become capable of fully building their own successors.

In that world, the pace of AI development is no longer primarily human-paced. It becomes compute-paced. Humans may still supervise, but supervision becomes meaningful only if humans can actually understand, verify, and intervene in what the systems are doing.

That is the central governance problem.

The question is no longer simply whether AI can generate useful code. The question is whether society can maintain a real control loop over systems that are increasingly involved in designing, testing, and improving themselves.

This is why Anthropic argues that the world may need the option to slow or temporarily pause frontier AI development if capabilities move faster than alignment research, public institutions, and safety mechanisms can handle.

That proposal is controversial. Critics will say it benefits incumbents. Others will say it is unrealistic without global coordination. Both criticisms are fair. A unilateral pause by one company would not solve the problem if competitors or governments continue racing ahead. A credible pause would require verification, trust, and international coordination the kind of infrastructure that usually takes years or decades to build.

But dismissing the concern would be a mistake.

The real story is not that AI has already escaped human control. It has not.

The real story is that the role of humans in AI development is becoming thinner, more strategic, and more fragile. We are moving from creators to supervisors, from builders to reviewers, from operators to governors.

That can be powerful if done well.

It can also become dangerous if the human layer turns into a rubber stamp.

The next phase of AI will not be defined only by bigger models. It will be defined by who controls the loop: humans using AI to build better systems, or AI systems increasingly building the next version of themselves while humans struggle to keep up.

Anthropic’s warning should not be read as panic.

It should be read as a signal.

The age of AI-assisted development has already arrived. The age of AI-directed development may be next. The question is whether our institutions, companies, and safety systems can evolve before the loop closes.


Written by Guy Yanpolskiy
GYG.agency by GYGlobal.net


Guy Yanpolskiy is the founder of GYG.agency and co-founder of GYGlobal Holding, working at the intersection of AI visibility, GEO/AEO, reputation architecture, PR, and international business events. He helps leaders, companies, and emerging brands become discoverable, trusted, and correctly interpreted by both humans and AI systems.