On 18 June 2026, the Anthropic Institute published an article that should give pause to anyone working in business, technology or consulting: “When AI builds itself” — in other words, when artificial intelligence starts building itself.
This is not science fiction. It is real internal data, published for the first time, describing what is happening today in Anthropic’s labs.
The numbers that change everything
The most striking figure: over 80% of the code merged into Anthropic’s codebase is already written by Claude. Until February 2025, this percentage was in the low single digits.
Productivity per engineer? 8 times higher in the second quarter of 2026 compared to 2024. Not because engineers are working more — but because Claude writes the code and they direct and review it.
But it’s not just code. Researchers working on open-ended and complex problems — deciding which experiment to run, interpreting results, choosing the next direction — already see Claude outperforming human choices in 64% of cases when it comes to deciding the next step in a research session.
The concept that changes the rules: recursive self-improvement
The article’s title refers to recursive self-improvement: the ability of an AI system to autonomously design and develop its own successor. Anthropic is clear: we are not there yet. But the signals indicate we are getting closer.
The duration of tasks an AI can complete autonomously doubles every four months. In April 2024, Claude could handle 4-minute tasks. A year later, tasks lasting an hour and a half. Today, 12-hour tasks. If the trend holds, by 2027 AI could be handling tasks that take weeks.
This is not a linear trend. It’s exponential. And exponential trends, by definition, always surprise.
What this means for businesses
For anyone still using AI as an “assistant for writing emails”, this is the moment to recalibrate expectations — upward.
What is emerging is not a tool that speeds up individual work. It is a structural shift in the relationship between human and artificial intelligence. Companies that integrate AI into their processes today — with secure tools, in controlled environments, with data guarantees — are building a competitive advantage that in two years will be almost impossible to recover for those who haven’t started.
Anthropic itself describes a near future in which every person in a company coordinates a “pyramid of agents” — AI systems that carry out operational work while the employee focuses on direction, judgement and strategy.
The question that matters
The question is not “will this technology work?” — the answer is yes, the data is already there. The question is: is your company already working to integrate it safely and in a structured way?
AIDeskPro was designed precisely for this: bringing the best generative AI models — Gemini, Claude, GPT — into a secure corporate environment, with EU data residency guarantees and zero data retention. Not a consumer tool, but an enterprise platform ready to face what is described in Anthropic’s article.
The future is here. It’s just waiting to be adopted.
Original source: When AI builds itself — Anthropic Institute, 18 June 2026.