Boris Cherny, the creator of Anthropic’s generative coding tool Claude Code, has admitted that letting AI write 100% of code is becoming ‘problematic’ for companies, even as he continues to champion AI’s transformative role in software development. According to a report by Business Insider, speaking at a fireside chat hosted by Scale AI, Cherny said companies are right to focus on return on investment (ROI) as AI token costs rise.He responded directly to concerns raised by Uber COO Andrew Macdonald, who recently questioned whether AI spending was delivering enough consumer-facing value. “ROI is absolutely the right framing because you don’t want to just think about cost… you spend something on it and you get something back,” Cherny said. He cautioned against restricting token usage too early, arguing that experimentation often leads to breakthrough ideas.
AI writing all code creates new bottlenecks
Cherny further acknowledged that measuring the impact of AI by the percentage of code written is no longer useful, since many engineers now allow AI to generate their code. “Once you get it to this point where engineers are just writing a lot of code, the bottleneck is going to be good ideas,” he said. He also warned that companies must focus on idea generation and innovation pipelines, rather than simply accelerating code output.
Anthropic’s cost controls
Cherny also stressed on the fact that Anthropic offers enterprise customers ways to manage token budgets, including per-seat cost controls and backend usage monitoring. He also noted that tokens are not free for Anthropic either, “Every token we use is a token we do not give to a customer, so there’s an opportunity cost.”
Boris Cherny feels days of AI prompts are over
After previously declaring that “software engineering is dead,” Cherny now says the era of manually writing AI prompts is ending. According to a recent report by Business Insider, Cherny argues that the future lies in loop engineering, a system in which AI agents generate and refine prompts themselves. For those unaware, loops are the recurring systems that guide AI agents without constant human input. For instance, a command like /goal can instruct an AI model to keep working until a task is complete, rather than requiring step-by-step prompts. Cherny explained: “It’s an agent that prompts Claude. I don’t write the prompt anymore. Claude writes the prompt, and now I’m talking to that new Claude that is coordinating.”While loops reduce human effort, they raise concerns about token budgets. Running multiple agents and sub-agents can quickly become expensive. Steinberger advised using longer intervals — hourly or daily — to reduce costs, while Osmani cautioned that sub-agents should only be used when a second opinion is worth the expense.
