The question sounds almost naive now.
If AI can generate apps, fix bugs, write functions, review pull requests, explain unfamiliar codebases, and work across files for hours at a time, why should anyone still learn to code?
It is an understandable question. The software world has spent the past year watching coding agents move from impressive demos into daily workflows. Developers are no longer only asking chatbots for help with syntax. They are delegating tasks, comparing outputs, supervising agents, and waking up to code that was written while they were away from the keyboard.
At the same time, coding is no longer only a developer conversation. Founders use AI tools to prototype products before hiring technical teams. Product managers use them to test ideas. Designers use them to make interfaces interactive. Operators use them to automate internal workflows. The ability to create software is spreading beyond the people who traditionally called themselves software engineers.
So the better question may not be whether people should still learn to code.
It is what “learning to code” now means.
The old answer was mostly about syntax, frameworks, and the discipline of building software line by line. That still matters. But AI is changing the center of gravity. The value is moving toward judgment: knowing what to build, how to describe it, what context the system needs, whether the output is good, and where the risks are hiding.
In other words, the future of coding may involve less typing.
But it will require more understanding.
Vibe coding to real work
That transformation was visible at SXSW this year during a session with Bolt.new and Anthropic. The conversation was not about the fantasy version of AI coding, where a person writes one sentence and a perfect product appears. It was about something more mature: how agentic coding tools are moving from playful prototypes into real company workflows.
Bolt became one of the clearest examples of the new AI coding wave because it made software creation feel immediate. A user could describe an app and quickly see something working on screen. That experience helped popularize the language of “vibe coding,” a phrase that captures both the magic and the danger of the moment.
But at SXSW, Bolt CEO Eric Simons described a more structured use case. The value of Bolt, he said, is often “rapid prototyping.” Not replacing the production codebase. Not letting every employee ship directly to customers. Instead, it gives teams a way to explore how a feature should look, feel, and behave before engineering time is committed.
That distinction matters.
In many companies, the bottleneck is not only writing code. It is alignment. Product teams, designers, executives, and customers often need to see and touch an idea before they can decide whether it is worth building. AI tools make that exploration faster. But the more serious opportunity is connecting those prototypes to the systems a company already uses.
Simons described a workflow where Bolt can work with a company’s actual design system, UI components, and API shapes, so a prototype is not just a disposable mockup. Once the experience is locked in, a coding agent such as Claude Code can help translate that work into something closer to the production environment.
The important part is the handoff. Non-engineers can shape ideas in a sandbox. Engineers still protect the production system. Between the two, the agent becomes a bridge.
That is a very different story from “AI replaces developers.” It is closer to this: AI changes who gets to participate in software work, and it changes what developers are needed for.
Context is becoming the new skill
This is also why the conversation has moved beyond prompt engineering.
A few years ago, much of the advice around AI tools focused on prompts: how to phrase instructions, how to structure requests, how to coax better answers out of the model. That skill still has some value, but it is becoming less central as models improve.
The stronger pattern now is context.
Anthropic’s spokesperson described “Skills” as a way to give agents useful knowledge they can call on when needed. Instead of forcing every instruction into one perfect prompt, teams can give agents access to the right documents, rules, examples, and workflows.
For companies, that is where the real work begins. A useful coding agent does not only need a request. It needs to understand the environment around the request: the design system, the codebase, the testing culture, the security expectations, the product logic, and the conventions that make one company’s software different from another’s.
Someone has to know what context matters.
Someone has to know whether the agent is using the right component, respecting the right constraint, or making a decision that will create problems later.
The person who understands the system becomes more valuable, not less.
The coding interview is already changing
The same evolution is starting to appear in hiring. One of the most interesting parts of the SXSW session was not about tools, but about talent.
Simons said that “some of the most amazing people” Bolt had hired in the past six months either did not pass the old technical interview or would only have passed after the company changed its process. What made them exceptional, he said, was that they were “using agentic tools to get the job done.”
That does not mean the technical interview is dead. It means the signal is changing.
Bolt now asks candidates what AI tools they use and how they use them. The point is not whether someone can name popular products. It is whether they have actually explored them deeply enough to build something meaningful.
As Simons put it: “Show us what you built, show us how it works. If it’s real, then okay, this person can do the work.”
That feels much closer to the reality of work now. If AI tools are part of the job, evaluating candidates as if those tools do not exist gives an incomplete picture. But the opposite is also true. If someone uses AI without understanding what it produced, that is not fluency. It is dependency.
Anthropic’s side of the conversation kept that tension intact. The company still uses at least one interview where candidates work without AI assistance, reading and writing some Python. But the goal is not to test obscure syntax tricks. It is to see whether someone can understand patterns, debug an agent, and reason through a system.
That may be the new balance: fluency with agents, plus enough technical depth to know when they are wrong.
So, should people still learn to code?
Yes. But not because everyone needs to become a traditional software engineer.
They should learn because software is increasingly the surface through which work gets done. Even when AI writes the first draft, people still need to understand what is being created. They need to know when something is fragile, when it is secure, when it is scalable, and when it only looks impressive in a demo.
Simons made the point directly. “Writing software is perhaps a smaller problem now,” he said. “But what about reviewing? How do we scale this?”
That may be the most honest version of the answer.
The easier it becomes to generate code, the more important it becomes to know whether that code should exist, whether it works, and whether it can be trusted.
This is why coding literacy does not disappear. It changes shape. It becomes less about memorizing syntax and more about understanding systems. Less about producing every line and more about directing, reviewing, testing, and improving what agents produce.
The year agentic coding starts to mature
The industry is already moving in that direction.
In June, AWS added OpenAI’s Codex coding agent to Amazon Bedrock, making it generally available to enterprise customers through AWS infrastructure and a pay-per-token model. It is a small but telling signal: agentic coding is becoming part of the cloud platforms and procurement channels where large companies actually adopt software.
Anthropic has continued to push Claude Code toward larger and more complex workflows. OpenAI has positioned Codex not only as a coding tool, but as a broader agent for knowledge work. The direction is clear: agents are moving from side experiments into the daily structure of how work gets done.
But the more interesting story is not that the tools are getting better. It is that companies are starting to design around them.
That means new workflows, new pricing models, new hiring signals, new management habits, and new expectations for employees. It also means a new divide: not between people who can code and people who cannot, but between people who can work effectively with intelligent systems and people who are still waiting for the dust to settle.
At SXSW, Bolt and Anthropic captured that moment of transition. The first wave of AI coding was euphoric, messy, and experimental. The next wave is more operational. It is about permissions, context, design systems, testing, review, and safe handoffs into production.
As Simons said near the end of the session, companies are beginning to adopt these tools “in real ways, not just exploratory toy ways.”
“This is the year of maturing,” he said.
That is less flashy than the promise of instant software. But it is much more important.
The future of coding is not simply that machines will write more of it. They will. The future is that more people will be able to participate in shaping software, while the people who understand software deeply will be responsible for making sure it works.
So yes, you should still learn to code.
Not because the world needs everyone to type every line by hand. But because the world is being rebuilt through software, and the people who understand how it works will be better equipped to guide the agents rebuilding it.
