Should You Still Learn to Code in the Age of AI Agents?
The 2025 Edition
TL;DR
Should you learn how to code in 2025?
Here’s the recap:
Demand: The need for Software has been growing exponentially since the 1940s, with no signs of slowing down. Software continues to eat the world.
Supply: LLMs will make existing engineers more productive and will lower the barrier of entry, allowing non-technical people to produce Software. It’s too early to tell if this increase in supply will be able to meet the ever-growing demand.
Adoption: AI coding practices are following an S-curve, with broader adoption of full automated agents potentially taking decades.
Technology: While improving rapidly, AI-generated code still faces severe limitations, making human oversight and validation crucial.
Common Language: Code remains the universal medium for collaboration between humans and machines. This could change in the future (new languages and tooling exclusive to machines could emerge), but we have no evidence for that.
Conclusion: Yes, this is an excellent time to learn to code. Coding is the literacy of this century. If you want to contribute and lead in this new industrial revolution, understanding the fabric, rules and elements of how it works will give you a massive advantage.
The World Is Fast and Noisy
Just a year ago, the answer to this question was a straightforward "Yes!" However, the rapid changes in the industry have made the answer less clear and now require deeper analysis.
In less than a year, Large Language Models (LLMs) have significantly improved in coding capabilities. AI agents have become the latest trend, and "vibe coding" is now a term recognized even by non-technical audiences.
Adding to this noise, several CEOs from major companies keep making bold, yet unsubstantiated claims, further muddying the waters.
“Over the last 10-15 years, almost everybody who sits on a stage like this would tell you that it is vital that your children learn computer science, everybody should learn how to program. In fact, it is almost exactly the opposite,”
- Jensen Huang (Nvidia CEO)
Navigating through all this change and noise is difficult, and predicting the future is impossible. However, we can:
Examine major economic trends (supply and demand)
Learn from past technology adoptions and disruptions
Let’s start by looking at the supply/demand.
Nearly Infinite Demand for Software
From a demand perspective, the need for Software has been insatiable since the 1940s, showing no sign of slowing down. Software has been eating the world for decades now, yet we’ve barely scratched the surface.
In practice, many real world workflows remain far from the digital efficiency we've been promised: siloed applications, clunky interfaces, reliance on emails and phone calls, physical queues, manual steps, and frequent data copying between systems.
Despite the rise of SaaS, low-code platforms, and digital transformations, these issues persist. In fact, the proliferation of systems has often exacerbated the problem.
Additionally, each new layer of Software opens up more possibilities and powerful abstractions, generating even more ideas, opportunities and demand. Over the past two decades, we've witnessed the emergence of app stores, reactive web apps, cloud computing, containers, Kubernetes and data lakes, leading to entirely new opportunities and career paths.
Since Software's inception in the 1940s, demand has grown exponentially, with no signs of abating. It's reasonable to expect this trajectory to continue for the next 30-50 years or more.
Constrained Supply
On the supply side, we're producing more Software than ever. This trend predates LLMs, with Software becoming integral and crucial to every facet of our lives.
However, a significant bottleneck remains: the availability and high cost of talent capable of producing high-quality, well-engineered Software. Many traditional companies struggle to afford, attract, and retain top engineering talent.
This is where LLMs and code-generation agents come into play:
They enable existing engineers to be more productive, hopefully producing more and better Software.
They lower the bar of entry, enabling more people (domain experts) to produce Software without years of experience and a formal engineering education.
The key question is: will LLMs and code-generation agents finally meet the world’s insatiable demand for Software? It is too early to say, but we see no meaningful signs of that.
Let’s explore what this means in more detail.
A New Spectrum of Developers
As we navigate the evolving landscape of Software development, we're witnessing the emergence of a diverse spectrum of developer profiles. These range from traditional manual coders to those embracing full automation. This spectrum isn't rigid; it's fluid, with individuals and teams adopting different approaches based on their specific contexts, projects and codebases.
Most importantly, we might see a mix of profiles working in the same team and contributing to the same codebase.
Manual Developers - These are seasoned developers who prefer to write code manually. They might be skeptical of AI-assisted tools or operate in specialized domains where precision, quality and control are essential.
Augmented Developers - These developers leverage AI tools like Cursor or Copilot to enhance their productivity. They use AI to generate boilerplate code, refactor existing code or explore new solutions, all while maintaining oversight and control over every single line of code produced.
Automated Developers with Human in the Loop (aka “Vibe Coding”) - Often domain experts with limited coding experience, these individuals use AI to generate code based on their specifications. They rely on iterative feedback, reviewing and refining the AI-generated code to meet their requirements.
Fully Automated Agents - Fully autonomous AI agents capable of generating complete Software solutions based on high-level specifications, with minimal to no human intervention. While promising and exciting, this level of automation is still in its infancy and not yet widely adopted for anything real.
As of April 2025, we see:
Manual developers remain prevalent, especially in industries requiring stringent compliance, quality and engineered precision.
Augmented developers are on the rise and becoming the norm. In a Github and Accenture joint study done in 2024, of 50,000 developers, 80 % activated Copilot licences and 67 % used it at least five days a week
Automated developers with human in the loop are increasingly common in rapid prototyping scenarios. However, the technology isn't yet mature enough for production-grade systems.
Fully automated development remains largely experimental. While there are promising developments, fully autonomous AI-generated Software solutions without human oversight are not yet commercially viable.
Adoption Speed
Contrary to what some CEOs might be wishing for, the adoption speed of this technology might not be as fast as they expect.
“I think we will be there in three to six months, where AI is writing 90% of the code. And then, in 12 months, we may be in a world where AI is writing essentially all of the code.”
- Dario Amodei (Anthropic CEO), March 11, 2025
As Kent Beck pointed out recently, the adoption of new technology always follows an S curve, whether this is for hybrid corn or for AI. This idea was also popularized in the famous book “Crossing the Chasm” by Geoffrey Moore.
What this means in practice is that:
AI coding practices will take a lot longer to propagate and be adopted than most of us think. Yes, yes - even in the tech industry.
Adoption is not homogeneous; we will see both early adopters jumping head-first into vibe coding their way into production with all the risks associated, and companies that will delay and resist the adoption of this technology for years to come.
This is a sociological factor that has nothing to do with how loud your CEO is shouting, but how we as humans transmit knowledge and success stories through peer-to-peer trust relationships.
Additionally, as of April 2025, the technology is still far from being safe to use without close human oversight and validation. While the future may belong to AI-native development, today's LLMs still hallucinate APIs, produce inconsistent code, struggle with complex debugging, and lack architectural understanding. They often generate overconfident but flawed solutions, especially dangerous in production.
These aren’t minor glitches; they fundamentally limit trust and require robust validation, experienced oversight, and process changes that most teams aren't ready for. Until engineering culture, tooling, and practices evolve to absorb these shifts, vibe coding will remain a powerful tool, but not yet the default way we build Software.
Code as The Common Language
One of the strongest arguments for learning to code is that it remains the common language among all types of developers. It's the lingua franca enabling humans and coding Agents to collaborate on the same codebase.
In likely scenarios where adoption is slower and technology maturation takes time, humans and machines will need to coexist and complement each other's abilities.
From this perspective, learning to code is akin to learning reading, writing, or math. Despite technological advancements, these foundational skills remain essential. Coding is becoming the new literacy of this century.
Looking further ahead, we might see new languages and tools designed exclusively for automated coding agents, but there's no evidence of that today.