Claude Is Writing Anthropic’s Code. The Human Role Is Changing
by Jaffa Levy
·
Anthropic has published one of the clearest signs yet that artificial intelligence has entered the machinery by which artificial intelligence is built.
In a paper from the Anthropic Institute titled When AI builds itself, the company says that, as of May 2026, more than 80 percent of the code merged into Anthropic’s codebase was authored by Claude. Before Claude Code launched in research preview in February 2025, the figure was in the low single digits. Anthropic also says the typical engineer was merging eight times as much code per day in the second quarter of 2026 as in 2024.
The claim is striking, but it needs to be read carefully.
Claude is not running Anthropic. It has not independently designed its own successor, chosen its own goals and merged its own improvements without human control. Anthropic is not claiming that full recursive self improvement has arrived. Its argument is more precise: AI is now accelerating the work of AI development itself, and the distance between assisted engineering and autonomous improvement may be narrowing faster than institutions can absorb.
The central distinction: Anthropic’s disclosure does not show that humans have disappeared from AI development. It shows that the human role is moving upward. Engineers are increasingly setting objectives, supervising agents, reviewing machine generated work and exercising judgement over systems that can now supply much of the method themselves.
The deeper shift is in the role of the human engineer.
Engineers are no longer merely writing code. They are setting objectives, supervising agents, reviewing machine generated work and exercising judgement over systems that can now supply much of the method themselves. The old model was simple: humans decided, humans wrote, humans tested, humans merged. The emerging model is different: humans decide, AI writes, AI tests, humans review, humans merge.
That is not the end of human engineering. It is the reorganisation of it.
The crucial distinction is between goal and method. Humans still supply the goal. Claude increasingly supplies the method. Humans still decide what matters, whether the answer is good enough, whether the result should be trusted and whether the change should enter the codebase.
When code generation was slow, the bottleneck was human implementation. When AI can produce code quickly, the bottleneck moves to judgement, review, testing, safety, architecture and governance. More output does not automatically mean better output. It means the human supervisor becomes more important, because one person can now approve or reject far more machine generated work than before.
The new risk: A good engineer with Claude may become far more productive. A weak engineer with Claude may become dangerous at scale. The tool multiplies judgement. It does not replace the need for it.
The broader implication is institutional. If AI can increasingly perform the method layer in software development, the same question will spread to research, finance, law, science and government. Who sets the goal? Who grants permission? Who verifies the result? Who carries liability when the machine acts quickly and the human review becomes thinner?
The discussion around Anthropic’s paper has understandably reached for large historical analogies: nuclear weapons, the Cuban missile crisis, the industrial revolution, the singularity. Some of that language is inflated. The more immediate change is quieter. It is happening inside workflows, repositories, testing systems and review processes.
First the machine writes suggestions. Then it writes functions. Then it handles tasks. Then it works for hours. Then the human stops watching every step and begins reviewing the outcome.
Autonomy enters institutions through usefulness before it enters them through formal authority.
Anthropic’s disclosure is therefore not proof that the machine has escaped. It is evidence that the location of human control is changing. The human has not vanished. But the human is no longer standing in the same place.
AI is not yet building itself alone.
But it is now helping build the systems that will build the next systems.
