Anthropic says Claude ports COBOL; IBM shares slide

The article opens as a personal recollection: the writer, once a CTO, would block out a full day at the office, write code in one burst, then hand it off to engineers for verification, integration, commits and maintenance. Back then, writing code meant typing every line by hand. Now, the writer argues, anyone can be a coder — not in the old sense, but by instructing AI in natural language and generating code in a language of choice. Early generated code was visibly poor, but quality has improved dramatically, and Anthropic and several other industry executives have said they expect a substantial share of their own code to be AI-generated in the future. The writer's first implication is that engineers are still needed: while writing code is a large part of their training, production code demands many hours of maintenance per hour spent generating it, and while AI agents will take on much of that maintenance, humans will supervise them — a shift captured by the phrase Humans On The Loop.

The second implication asks whether software vendors are still needed. The logic is clear: if a company can develop, launch, operate and maintain its own service applications using a team of AI agents plus human oversight, why pay a vendor? Anthropic's COBOL announcement is the concrete example here, and the investor reaction — fear that IBM's long-standing strength in truly understanding such systems becomes unnecessary — is what moved IBM shares. The writer says time will tell, and points out that vendors supply much invisible value beyond software: battle-tested features and an unseen expert team whose knowledge comes from customers' workloads, not just one's own. How much that is worth, and whether every customer can obtain it individually or only certain segments can, is also for time to tell.

On productivity, the article notes a report that Accenture tracks employees' AI usage as a promotion criterion. It says leaders at every level are wrestling with how to compare human productivity with AI-augmented human productivity, and that it is easy to say people should use AI or that humans and AI together outperform either alone. The writer has written about an "AI force multiplier", and says the threshold question is how productive an AI-using person must be to justify their salary plus token costs. Many companies are freezing hiring and making teams prove a new role cannot be replaced by AI, and the writer calls for more quantitative, measurable, ideally industry-standard methods to assess ROI. A further implication is code proliferation: as agents grow adept at maintaining their own code, generating new code per task may become more cost-efficient and stable than modifying existing code, potentially leaving each function on its own codebase or agent. That would overturn many software lifecycle management practices, simplifying some work while dramatically complicating others. Finally, on security, Anthropic recently mentioned the number of distillation attacks its LLM Claude received — attacks that repeatedly query an AI to understand and replicate it. As more companies embed externally exposed AI into products, it is unclear whether the existing practices remain the best way to handle this. The writer's closing recommendation is that every company should have its own AI software strategy, that those just starting should consider an AI-native strategy leading into a software strategy, and that strategy should be kept current as new AI technologies and best practices keep coming.