At Anthropic, an AI lab creating some of the most cutting-edge models in the world, engineers are outsourcing writing of the code that drives their products to AI instead of doing it themselves. Boris Cherny, the head of Anthropic’s Claude Code, has declared that he hasn’t written any code in over two months.
Cherny said in a post on X that Opus 4.5 and Claude Code from Anthropic now write all of his code. He claims that “pretty much 100%” of the code in the rest of the corporation is also produced by AI.
In response to AI researcher Andrej Karpathy, Cherny commented in a post on X, “For me personally, it has been 100% for two+ months now, I don’t even make small edits by hand.” “I shipped twenty-two pull requests (PRs) yesterday and twenty-seven the day before, all of which were entirely written by Claude.”
The comments are in line with prior statements made by Dario Amodei, CEO of Anthropic, during the World Economic Forum earlier this month. He mentioned that some of his company’s engineers no longer write code directly, instead depending on AI models to do so while they concentrate on editing. Amodei said at Davos that the industry might be six to twelve months away from AI managing the majority or all of software engineering tasks from beginning to end.
Cherny is not the only well-known engineer to declare that they have mostly given up on hand coding. In a post on X, Roon, a well-known pseudonymous account created by an OpenAI researcher, also stated that he no longer develops any of his own code. The user responded, “100%, I don’t write code anymore,” when asked what proportion of his coding is handled by AI models. He said in a another post: “It was never good at programming. I’m relieved that it’s finished since it was a necessary pain for anyone who wished to control computers to accomplish beneficial tasks.
Although people in the business have an incentive to promote their own tools, there is increasing agreement that the emergence of AI coding tools has already had a significant impact on the sector.
Himanshu Tyagi, co-founder of the open-source AGI startup Sentient, told that “the entire way people build software has changed; software is not like what it used to be.” AI will write a significant portion of the code that is deployed during the next ten years. [Anthropic’s] Claude Code is the ground-breaking product that has enabled this to occur.
But outside of the top AI labs, many software companies report much lower numbers for AI-generated code. In April 2025, Microsoft CEO Satya Nadella, for example, stated that over 30% of the company’s code was generated by AI. A comparable number has been released by Salesforce. According to a study on GitHub Python functions that was published in the journal Science earlier this month, over 29% of them are now AI-written in the United States, with smaller numbers in other regions.
Additionally, according to Cherny, AI now writes all of his code, although an Anthropic representative stated that the company-wide percentage is closer to 70% to 90%. About 90% of Claude Code’s code is created by Claude Code itself.
According to Cherny, these numbers will only rise further, and other businesses will soon begin to generate AI code at comparable levels. He stated, “I believe that in the upcoming months, the majority of the industry will see similar stats—it will take more time for some vs. others.” “After that, we’ll see comparable statistics for computer work that isn’t coding.”
In recent years, Anthropic’s tools have gained popularity among software programmers. However, the release of Claude Code has resonated with both coders and non-coders, resulting in a viral moment for the company unlike anything witnessed since ChatGPT’s launch. After consumers pointed out that Claude Code was more of a general-purpose AI agent, Anthropic developed a version of the product for non-coders, introducing Cowork, a file management agent that functions similarly to the coding software. Cherny stated that his team constructed Cowork in about a week and a half, mostly utilizing Claude Code itself.
Cherny claims that the tool was causing a stir within their own organization even prior to the public uproar.
Approximately a year ago…In an interview this week, Cherny stated, “We thought the model was strong enough that we could use it for a different kind of coding. We started to try it out internally, and it just took off.” Because Claude handles practically all of the hard labor and I get to be creative, I have never enjoyed my job as much on a daily basis as I do now. I have time to consider what I want to construct next.
In addition to coding, Cherny claimed to utilize Claude Code for other administrative duties, such as project management duties like sending out Slack messages to team members when they fail to update shared spreadsheets.
He remarked, “Engineers simply feel unshackled because they don’t have to work on all the tedious stuff anymore.”
A reckoning for the software industry
The rise of AI-generated code has had a tremendous impact on the software industry. Many major tech businesses have been forthright about the fact that AI models write a big portion of their code. However, the automation of much of the coding process has generated concerns about the future of software engineering roles, particularly entry-level ones that have historically functioned as training grounds for the industry.
By enabling people with little to no technical expertise to create things by interacting with AI systems in plain language, tech companies contend that the quick uptake of AI coding tools like Claude Code and GitHub Copilot will democratize coding. The number of open positions for entry-level software engineers has decreased as the amount of code created by generative AI has increased, while it is not clear that the two are causally related as there are other factors influencing a job’s decline.
The change is already altering Anthropic’s hiring practices. Cherny added that because many traditional programming abilities are less useful when AI handles implementation details, his organization now hires a lot of generalists rather than specialists.
Cherny wrote, “Not all of the things people learned in the past translate to coding with LLMs.” “The model can complete the details.”
Cherny noted that the technology is still in its infancy, but he also highlighted the productivity increases and creative freedom that AI coding tools offer. Karpathy’s review states that models may overcomplicate code, create dead code, and make “subtle conceptual errors.” Despite the drawbacks, Cherny and other engineers are optimistic that the quality of AI-generated code will only become better.
