Anthropic CEO: engineering will be automated

Dario Amodei did not just say artificial intelligence will help engineers write code. In January 2026 at the World Economic Forum in Davos, he said models might be 6 to 12 months away from doing “most, maybe all” of what software engineers do end to end. (entrepreneur.com) That landed differently because Amodei runs Anthropic, the company behind Claude. When the person selling one of the strongest coding systems says the job could compress fast, people hear both a forecast and a sales pitch at the same time. (cfr.org) (anthropic.com) The fight is partly over one phrase: “software engineering.” Writing code is only one slice of that job, like laying bricks on a construction site where someone still has to choose the land, read the blueprint, inspect the work, and decide what gets built next. (openai.com) Benchmarks usually test a narrower task. SWE-bench Verified, a widely cited test OpenAI described in August 2024, gives an artificial intelligence agent a real GitHub issue, a code repository, and a hidden test suite, then checks whether the patch fixes the bug without breaking other features. (openai.com) Anthropic’s own product page now says Claude Code reads a codebase, edits files across that codebase, runs tests, and can deliver committed code. The same page says a majority of Anthropic’s own code is now written by Claude Code, with engineers shifting toward architecture and orchestration. (anthropic.com) Anthropic’s research gives a clue to what is already changing inside the work. In a study of 500,000 coding-related interactions published on April 28, 2025, Anthropic said 79% of Claude Code conversations looked like automation rather than collaboration, compared with 49% on the general Claude chat product. (anthropic.com) That same report found web-development languages like JavaScript and HyperText Markup Language were the most common in the dataset, and user interface work was among the top uses. That points to a near-term squeeze on the parts of engineering that are repetitive, visual, and easy to test. (anthropic.com) The harder parts are the messy ones around the code. Big companies still have old systems, compliance rules, security reviews, vague product requests, and customers who change their minds halfway through a release, and those are not captured cleanly by a benchmark patch test. (openai.com) (anthropic.com) So the most literal reading of Amodei’s claim is not “every engineer disappears by early 2027.” It is closer to “the machine may soon do the typing, testing, and first draft of whole features, while humans move up a layer into judgment, review, and system design.” (entrepreneur.com) (anthropic.com) That is why his advice to engineers has an edge to it. If Claude Code can already refactor across many files, run continuous integration checks, and fix failures automatically, then the engineers who know how to supervise agents will look more valuable than the engineers who only know how to write code line by line. (anthropic.com) The real argument is no longer whether artificial intelligence will touch software jobs. The real argument is whether “fully automated” means one model can replace a whole engineering team, or whether it means one engineer can suddenly do the work that used to require five. (anthropic.com) (entrepreneur.com)