Anthropic and peers expect most code to be AI-written

The author's argument begins with his own experience: as a young CTO, writing code meant typing it line by line and handing it to engineers to verify, integrate, commit and maintain. That is no longer the only path. AI now lets anyone direct code generation in natural language, and while early generated code was visibly poor, the author says quality has improved dramatically. On that basis, he reports that Anthropic and other industry executives expect a substantial share of their code to be AI-generated in the future.

Much of the piece works through the consequences in sequence. Engineers remain necessary because production code demands far more hours of maintenance than generation, though AI agents are expected to absorb much of that work under human supervision — the direction the industry describes as Humans On The Loop. The same logic drives the debate over whether SaaS vendors could be displaced if companies can build and run their own service applications, an argument sharpened by Anthropic's announcement that Claude can port COBOL code and by the resulting slide in IBM shares. The author cautions that time will tell, noting vendors also supply less visible value such as field-tested features and expertise accumulated across many customers' workloads.

The remainder of the piece turns on measurement and risk. Leaders are described as unsure how to compare human productivity with AI-assisted productivity, and firms are reportedly freezing hiring and requiring teams to prove a role cannot be replaced by AI — a gap for which the author calls for more quantitative and ideally industry-standard ROI methods. He also raises the prospect of code proliferation, where generating fresh code per task becomes cheaper and more stable than modifying existing code, upending software lifecycle practices, and points to Anthropic's disclosure of distillation attacks on Claude as evidence that security review has not caught up with AI-generated code.