- by x32x01 ||
DHH, the creator of Ruby on Rails, recently made a striking statement on stage at Rails World in Austin: developers may no longer need to write code by hand in the traditional way. 😳
According to DHH, the approach at 37signals is increasingly simple: AI agents write the code, while developers step in when something goes wrong.
That does not mean programmers are becoming useless. His argument is more specific: typing the code yourself may no longer be the most valuable part of software development.
Instead of a developer manually writing every line, the AI agent handles much of the implementation. The human developer then reviews the result, identifies problems, and makes corrections when necessary.
DHH reportedly described this as a major shift from the way software has traditionally been developed.
The interesting part is that this position is different from comments he made less than a year earlier, when he said he still preferred writing code himself and did not consider AI capable enough.
Now, his view appears to have changed significantly.
The AI may write the code, but the human still needs to understand the code well enough to judge it.
And that distinction is probably the most important part of DHH's argument.
According to his comments, the new backend was written in Rust using AI agents.
What's interesting is that DHH has said he does not particularly like Rust.
His reasoning was essentially that if agents write the Rust code, he does not necessarily have to spend his time reading every line of it himself.
He also claimed that the new version uses approximately 99% less CPU and 95% less memory, while noting that these numbers are approximate.
Those figures should therefore be treated as DHH's own reported estimates rather than independently verified benchmarks.
The bigger change is the role of the programmer.
A developer may spend less time physically typing code and more time:
An AI agent can generate code extremely quickly, but speed does not automatically mean correctness.
Imagine a beginner asks an AI agent to build an authentication system.
The agent produces thousands of lines of code in seconds.
How does the beginner know whether the implementation has a serious security vulnerability?
How do they know whether the database design will become a performance problem?
How do they know whether the agent misunderstood an important requirement?
If you do not understand the underlying technology, reviewing AI-generated code becomes extremely difficult.
You may be able to ask an AI to write code.
But knowing whether the code is actually good is a different skill.
An agent can generate an implementation incredibly quickly.
But it does not necessarily know when it should stop.
It can make assumptions.
It can choose an inappropriate architecture.
It can introduce a security problem.
It can solve the wrong problem while producing perfectly valid-looking code.
The human developer is still the one who needs to recognize those situations.
That's why learning programming can become even more important in an AI-assisted workflow, not necessarily less important.
DHH's argument, as described here, is not that programming knowledge has become worthless.
The more useful interpretation is that manual code typing may become less important while programming knowledge remains extremely valuable.
If AI agents increasingly write the implementation, developers need to become better at understanding what they are looking at.
You should still learn:
You need to understand enough to tell the AI when it is wrong.
AI agents are changing that workflow.
The future described by DHH looks more like this:
Human → defines the problem → AI agent → generates the implementation → Human → reviews and validates the result.
That is a very different relationship with software development.
The programmer becomes less of a person who manually produces every line and more of a person who directs, reviews, validates, and takes responsibility for the software.
Whether this becomes the dominant development model remains an open question.
But one thing is already clear: AI coding agents are changing what it means to write software.
The valuable skill may increasingly be knowing: "Is this code actually correct?"
That requires experience.
It requires understanding how software works.
And it requires enough technical knowledge to recognize when an AI-generated solution looks reasonable but is actually wrong.
So if you're learning programming today, don't learn only how to make an AI generate code.
Learn why the code works.
Learn how to break it.
Learn how to test it.
Learn how to secure it.
Learn how to optimize it.
And most importantly, learn how to recognize when the AI is confidently making a mistake.
The keyboard may become less important.
Your ability to think like a programmer won't. 💻
According to DHH, the approach at 37signals is increasingly simple: AI agents write the code, while developers step in when something goes wrong.
That does not mean programmers are becoming useless. His argument is more specific: typing the code yourself may no longer be the most valuable part of software development.
What Changed at 37signals? 🤖
DHH said that 37signals has moved toward an agent-driven development workflow.Instead of a developer manually writing every line, the AI agent handles much of the implementation. The human developer then reviews the result, identifies problems, and makes corrections when necessary.
DHH reportedly described this as a major shift from the way software has traditionally been developed.
The interesting part is that this position is different from comments he made less than a year earlier, when he said he still preferred writing code himself and did not consider AI capable enough.
Now, his view appears to have changed significantly.
AI Agents Are Writing the Code
The basic workflow can be summarized like this:- The developer describes what needs to be built.
- The AI agent generates the implementation.
- The developer reviews the generated code.
- If the agent makes a mistake, the developer fixes it or asks the agent to correct it.
- The developer remains responsible for deciding whether the result is actually correct.
The AI may write the code, but the human still needs to understand the code well enough to judge it.
And that distinction is probably the most important part of DHH's argument.
HEY's Backend Was Rebuilt With Rust
One of the more surprising examples mentioned by DHH is the backend of HEY.According to his comments, the new backend was written in Rust using AI agents.
What's interesting is that DHH has said he does not particularly like Rust.
His reasoning was essentially that if agents write the Rust code, he does not necessarily have to spend his time reading every line of it himself.
He also claimed that the new version uses approximately 99% less CPU and 95% less memory, while noting that these numbers are approximate.
Those figures should therefore be treated as DHH's own reported estimates rather than independently verified benchmarks.
Does This Mean Programmers Are Becoming Useless? 🤔
Not exactly.The bigger change is the role of the programmer.
A developer may spend less time physically typing code and more time:
- Understanding the problem
- Designing the system
- Reviewing AI-generated code
- Finding bugs
- Checking security issues
- Evaluating performance
- Testing the implementation
- Making architectural decisions
- Knowing when the AI-generated solution is wrong
An AI agent can generate code extremely quickly, but speed does not automatically mean correctness.
The Real Problem for Beginners
This is where things get more complicated.Imagine a beginner asks an AI agent to build an authentication system.
The agent produces thousands of lines of code in seconds.
How does the beginner know whether the implementation has a serious security vulnerability?
How do they know whether the database design will become a performance problem?
How do they know whether the agent misunderstood an important requirement?
If you do not understand the underlying technology, reviewing AI-generated code becomes extremely difficult.
You may be able to ask an AI to write code.
But knowing whether the code is actually good is a different skill.
AI Can Write Faster Than You Can Review ⚡
One of the biggest changes brought by coding agents is the difference between generation speed and human understanding.An agent can generate an implementation incredibly quickly.
But it does not necessarily know when it should stop.
It can make assumptions.
It can choose an inappropriate architecture.
It can introduce a security problem.
It can solve the wrong problem while producing perfectly valid-looking code.
The human developer is still the one who needs to recognize those situations.
That's why learning programming can become even more important in an AI-assisted workflow, not necessarily less important.
Should You Stop Learning to Code?
No.DHH's argument, as described here, is not that programming knowledge has become worthless.
The more useful interpretation is that manual code typing may become less important while programming knowledge remains extremely valuable.
If AI agents increasingly write the implementation, developers need to become better at understanding what they are looking at.
You should still learn:
- Programming fundamentals
- Data structures and algorithms
- Databases
- Networking
- Operating systems
- Security
- Software architecture
- Testing
- Debugging
- Performance
You need to understand enough to tell the AI when it is wrong.
The Programmer's Role Is Changing
For decades, being a programmer often meant spending a large part of the day writing code.AI agents are changing that workflow.
The future described by DHH looks more like this:
Human → defines the problem → AI agent → generates the implementation → Human → reviews and validates the result.
That is a very different relationship with software development.
The programmer becomes less of a person who manually produces every line and more of a person who directs, reviews, validates, and takes responsibility for the software.
Whether this becomes the dominant development model remains an open question.
But one thing is already clear: AI coding agents are changing what it means to write software.
The Skill That May Matter Most 🧠
If AI can generate code for almost anyone, then the ability to simply produce code becomes less distinctive.The valuable skill may increasingly be knowing: "Is this code actually correct?"
That requires experience.
It requires understanding how software works.
And it requires enough technical knowledge to recognize when an AI-generated solution looks reasonable but is actually wrong.
So if you're learning programming today, don't learn only how to make an AI generate code.
Learn why the code works.
Learn how to break it.
Learn how to test it.
Learn how to secure it.
Learn how to optimize it.
And most importantly, learn how to recognize when the AI is confidently making a mistake.
The keyboard may become less important.
Your ability to think like a programmer won't. 💻