On 4 June, Anthropic published a report titled When AI builds itself, co-authored by the Anthropic Institute's Marina Favaro and company co-founder Jack Clark, arguing that the industry should prepare a mechanism to slow or temporarily halt frontier AI development. The hook that drew most of the attention was a single disclosure: Claude, Anthropic's own model family, now writes more than 80 percent of the code merged into the company's production systems. In other words, the firm building the model is already leaning heavily on the model to build the next one.
When the tool starts building the toolmaker
The report's central worry is recursive self improvement, a point at which an AI system can design and develop its successor with limited human involvement. Clark put the odds at roughly 60 percent that AI could train its own successor by the end of 2028. Anthropic frames this not as science fiction but as a near term engineering trajectory, and offers the 80 percent code figure as evidence that the curve is already bending. For workers, the unsettling implication is that even the highly paid engineering roles often assumed to be safe from automation sit inside the loop being automated.
A conditional, not a unilateral, pause
The proposal is carefully hedged. Anthropic is not promising to stop on its own. It says it would only slow down if several well resourced labs across multiple countries agreed to halt under identical conditions, if compliance could be verified by independent systems, and if governments and civil society were meaningfully involved. Critics note the obvious tension: a leading lab calling for a pause that depends on rivals agreeing first can look like a way to claim the moral high ground without surrendering ground in the race. Legal commentators have flagged that the plan leans on governance machinery that does not yet exist.
What it means for the people who build AI
For the future of work debate, the report quietly reframes the conversation. Most coverage of AI and jobs focuses on clerical, administrative and entry level roles being squeezed. Anthropic's disclosure points up the ladder, suggesting software engineering itself is being reshaped from the inside. If a frontier lab can run with models writing the bulk of its code, the question facing every other employer is how soon a similar ratio reaches their own teams, and whether the senior engineers who review and steer that output become more valuable or simply fewer in number.
Whatever one makes of the pause proposal, the more durable signal may be the candour about how AI is now made. A company explaining that its product builds most of itself is making a claim about the pace of change that no labour market forecast has fully priced in. The policy debate about slowing frontier development will run for years. The workforce question it raises, about what is left for humans to do when the tools start building the tools, is already live.
