Boris Cherny, the creator of Anthropic's Claude Code, has publicly acknowledged that letting artificial intelligence write 100% of a company's code is becoming "problematic," even as he continues to champion AI's role in software development. Speaking at a fireside chat hosted by Scale AI, Cherny addressed growing concerns about return on investment (ROI) as companies grapple with rising AI token costs and the need to justify spending on generative coding tools.

Why Is 100% AI-Generated Code Becoming a Problem?

Cherny's admission marks a significant shift in how industry leaders are thinking about AI-assisted development. He explained that once engineers reach a point where AI is generating most of their code, the real bottleneck shifts from code output to something far more fundamental: good ideas. "Once you get it to this point where engineers are just writing a lot of code, the bottleneck is going to be good ideas," Cherny stated.

This insight challenges the earlier narrative that AI coding tools would simply accelerate development by automating routine programming tasks. Instead, Cherny argues that companies must now focus on idea generation and innovation pipelines rather than simply trying to maximize the volume of code their AI systems produce. The shift reflects a maturing understanding of where AI actually adds value in software development.

Cherny also responded directly to concerns raised by Uber's Chief Operating Officer Andrew Macdonald, who recently questioned whether AI spending was delivering enough consumer-facing value. Cherny acknowledged that ROI is "absolutely the right framing" for evaluating AI investments, noting that companies should think carefully about what they spend on AI tools and what tangible returns they receive.

Cherny

How Should Companies Manage AI Coding Costs and Token Budgets?

• Per-Seat Cost Controls: Anthropic offers enterprise customers ways to manage token budgets through per-seat cost controls, allowing organizations to set spending limits on a per-user basis.
• Backend Usage Monitoring: Companies can track and monitor how tokens are being consumed across their organization, providing visibility into where AI spending is concentrated.
• Cautious Experimentation: Cherny cautioned against restricting token usage too early, arguing that experimentation often leads to breakthrough ideas, but this must be balanced against cost concerns.
• Opportunity Cost Awareness: Anthropic itself faces opportunity costs with every token used, since "every token we use is a token we do not give to a customer," Cherny noted, emphasizing that token scarcity is a real constraint.

The tension between experimentation and cost control reflects a broader challenge facing enterprises adopting AI coding tools. Companies want to explore the full potential of these systems, but they also need to ensure that spending translates into measurable business value.

What Does the Future of AI Coding Look Like Beyond Manual Prompts?

Cherny has also declared that the era of manually writing AI prompts is ending. Instead, he envisions a future centered on "loop engineering," a system in which AI agents generate and refine their own prompts without constant human intervention.

In this model, commands like "/goal" can instruct an AI model to keep working on a task until it is complete, rather than requiring developers to write step-by-step prompts for each action. Cherny explained the concept this way: "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".

Cherny

"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," explained Boris Cherny.

Boris Cherny, Creator of Claude Code at Anthropic

While loops reduce the human effort required to guide AI systems, they introduce their own cost challenges. Running multiple agents and sub-agents can quickly become expensive, since each agent consumes tokens. This creates a new optimization problem for companies: determining when the value of having a second AI agent review or refine work justifies the additional token consumption.

The shift toward loop engineering represents a fundamental change in how developers will interact with AI coding tools. Rather than treating AI as a tool that responds to specific commands, developers will increasingly work with autonomous AI systems that manage their own workflows and decision-making processes. This evolution raises important questions about oversight, debugging, and the role of human judgment in the development process.

Cherny's candid acknowledgment that AI-generated code at scale creates new problems, rather than solving all of them, suggests that the industry is moving past the initial hype cycle around AI coding tools. The focus is shifting from "how much code can AI write" to "how can we use AI most effectively to improve software development outcomes," a more nuanced and realistic assessment of where these tools actually add value.