Ruby on Rails Creator Says His Engineers Will Now Let AI Write the Code

One of the world's most influential software developers has declared that writing code manually should no longer be the default way his engineers build software.

David Heinemeier Hansson, widely known as DHH and the creator of the Ruby on Rails web framework, says his company 37signals has adopted an AI-first approach in which coding agents will handle the initial work of producing code.

Speaking during his Rails World 2026 keynote, Hansson said manually writing code had effectively become an exceptional situation at the company rather than the normal development process. Engineers are expected to direct AI systems, review what they produce and intervene manually when necessary.

The shift is significant because Hansson has spent more than two decades writing software and shaping how web applications are built. Ruby on Rails, which he created in the early 2000s, became one of the most influential web development frameworks and has been used to build major internet businesses.

His latest position reflects how rapidly AI coding agents are changing software engineering.

Tools such as Claude Code, Codex and Cursor can now work across entire software projects rather than simply suggesting individual lines of code. Developers can describe a feature, ask an agent to inspect an existing codebase, generate files, run tests, identify errors and repeatedly modify the implementation.

That increasingly changes the developer's job from manually producing every line to directing and supervising systems capable of performing substantial portions of the implementation.

At 37signals, Hansson says engineers can still write code themselves when an AI agent fails to produce the required result. The difference is that such intervention is now treated as an exception. Once developers determine why the AI failed, they should improve the instructions or development process so the agent can handle similar work in the future.

His argument therefore goes beyond simply using an AI assistant alongside a conventional programming workflow. It suggests restructuring software development around agents, with humans increasingly responsible for architecture, requirements, judgement, verification and quality control.

Hansson has also pointed to his own output as evidence of how dramatically his development process has changed. During his Rails World presentation, he said he produced about 150,000 lines of code in August 2026, compared with a historical average of roughly 30,000 lines of production Ruby per year.

Lines of code are not necessarily a measure of software quality or developer productivity, but the difference illustrates the volume of implementation that AI systems can potentially generate when placed inside an agent-driven workflow.

The transition is also creating debate within the software industry.

Some developers see coding agents as an enormous productivity boost that removes repetitive implementation work and allows engineers to concentrate on architecture, product decisions and difficult technical problems. Others worry that excessive dependence on generated code could weaken developers' understanding of the systems they are responsible for maintaining.

Those concerns become particularly important when AI-generated software enters production. Generated code can still contain security vulnerabilities, inefficient logic, incorrect assumptions and subtle bugs, meaning human review, automated testing and quality assurance remain important even when humans are no longer typing most of the implementation themselves.

A recent Business Insider survey of more than 150 software engineers illustrated the divide. More than 45 per cent said they believed they were keeping up with AI tools, while 17.5 per cent said they were avoiding AI code-editing tools altogether.

Other engineers have raised a different concern: that software development could become less intellectually satisfying if programmers spend increasing amounts of time prompting, reviewing and correcting machines rather than solving implementation problems themselves.

A viral account from a software engineer recently described an AI-heavy development environment in which agents generated specifications, tests and other work while employees increasingly acted as supervisors. The account triggered wider discussion about whether productivity gains could come at the expense of engineers developing deep technical understanding.

Hansson is taking a substantially more optimistic position.

Rather than asking whether developers should resist AI coding, he argues that the important question is how engineers can extract the greatest benefit from increasingly capable agents.

That does not necessarily mean programming knowledge becomes irrelevant. Someone still needs to understand whether an application's architecture makes sense, whether the generated implementation meets requirements, whether security controls are adequate and whether the finished system actually solves the intended problem.

What may be changing is where that knowledge is applied.

Software engineers have historically spent much of their working time translating requirements into syntax and implementation. In an agent-driven model, more of that effort could move towards describing systems, designing architecture, establishing constraints, evaluating output and deciding what should be built.

For junior developers, the change could be particularly significant. Writing software manually has traditionally been one of the primary ways programmers learn how systems work. If AI increasingly performs that implementation from the beginning of a developer's career, companies and educators may need new ways of ensuring engineers still acquire the underlying knowledge required to identify when generated software is wrong.

Hansson's declaration does not establish that handwritten programming is disappearing across the software industry. Many organisations continue to use AI selectively, while others restrict its use because of security, intellectual property, reliability or regulatory concerns.

But the decision by the creator of one of the world's best-known programming frameworks to make AI-generated code the default at his own software company provides another indication of how quickly the profession is changing.

The question facing software engineers may increasingly be less about whether AI can write code and more about what humans should concentrate on once it can.