The New Age of Development

I have been working in tech for almost 20 years. Without a doubt, the shift we see in the software industry is unparalleled.

AI, which is how we have collectively decided to call LLMs, went from a trillion dollar party trick to something that can be actually useful, and the impact could not have been greater, not only due to its practical implications but also to a complex network of extreme hype, fear mongering, and massive hardware and software speculation.

With rapid iterations and very confusing marketing strategies, it became very hard even for the most critical person to assess whether the claims were true or just exaggeration in order to gain more media cycles. To make matters worse, we are facing completely new phenomena, like AI psychosis¹, which is not only real but very common in the tech bubble, especially amongst influencers apparently.

On top of that, we saw major, profitable corporations doing massive layoffs claiming to be due to very successful AI implementations, when in fact they had many other reasons, rapid inflation of their stock prices being one of the most common, so they could later quietly re-hire almost everyone. All under the false pretense of AI advancements. There is even a new term coined for that: AI washing².

Under all of this, I try to be as pragmatic as possible to assess the state of software development in 2026, and what I see is confusing, to say the least.

We have very polarizing positions, with pro and anti AI. For the sake of briefness, I will not enter into the debate of how moral or ethical the training of these models is; that is an entirely different question.

Pro

For those who preach the AI gospel, AI is the solution for every problem, and those who are there to hand code are coping or living in the past and will be "left behind".

It is indeed mesmerizing to see the agents spit out a working solution from a half-intelligible prompt. It looks like magic, and it is a technical marvel that this is even possible.

And that is one of the hardest parts of this discussion, because the measurement of productivity in software development is, and always was, elusive and practically impossible to be accurate. We created many different ways to do it along the way: story points, hours, complexity, etc. But the reality is that what dictates the productivity of a software product goes way beyond what the developer is able to put into code.

Even though AI infiltrated every single aspect of corporate life, the speed of release in real world scenarios and especially quality have not improved at the same pace as the AI models have.

An optimist might say that is because the companies do not embrace AI as they should, and that will be the argument of every single consultant from now on. Just like no one did Agile correctly, apparently no one will be able to do AI correctly.

Anti

On the other side of the fence we have the developers who are against: they avoid using it, judge those who do, and swear by the quality of the hand crafted solutions.

Claims of multiple X's of productivity, or a single developer that equals the output of whole departments, it's a dime a dozen online. But we have studies being published, that show the productivity gains are either negligible or even negative in some cases³.

Their concerns go deeper than productivity numbers though. A more fundamental worry is the erosion of understanding. When code is generated rather than written, the developer becomes a curator instead of a creator reviewing, accepting, and rejecting output without ever building the mental model that comes from working through a problem line by line. Debugging skills atrophy. System-level intuition never forms. Over time, the gap between what the code does and what the developer understands widens, and that gap is where catastrophic failures hide.

There is also the question of taste. Hand-crafted code carries decisions intent about naming, structure, boundaries, future extensibility that no prompt can fully capture. AI tends toward the average: the most statistically likely arrangement of tokens, which is rarely the most elegant or maintainable. The result is codebases that work today and are opaque tomorrow, full of patterns no human would chose, stitched together by statistics rather than design.

Critics also point to the "vibe coding" phenomenon: accepting output that looks correct without understanding it deeply enough to know whether it is correct. In safety-critical systems, regulated industries, or any code that outlives its original author, this is not a productivity gain it is a liability being deferred.

None of this means the anti side is right about the future. But their skepticism is not just Luddism. It is a bet that the long-term cost of code written without understanding will outweigh the short-term gain of code written faster.

Emerging Challenges

Whether you like it or not, AI seems to be here to stay; we cannot put the genie back in the bottle. And with it, a whole new host of challenges come into light.

Code Review

The most obvious for development teams is the bottleneck that the code reviewing step became, which, by the way, always had problems of its own, way before ChatGPT.

LLMs can produce an immense amount of code in a very short period of time, which leads to many more lines of code and MR to be reviewed. Reviewing can be as mentally demanding as, if not more than, writing the code, which means that either you glaze over, pretend you fully understand and approve if every step of the pipeline is green, or you use another LLM to help you out. Either way does not seem to be ideal; several layers of artificial information in between the developer and the actual intent of the code undermine the quality of the end result. And we have seen this in action multiple times with large, well funded corporations in the past year⁴.

This problem space is still pretty fresh and we don't seem to have a good enough solution for it.

Professional pipeline

Up until 2024, junior developers at the entry point of the software industry would write repetitive, boilerplate, or mundane tasks. Which is good, because that would slowly build their knowledge on standards, processes, syntax, patterns, etc. Now, this is pretty much done by agents, and without the proper incentive or even understanding of how dangerous this could be to the industry in the mid-long term, companies simply don't hire juniors anymore⁵.

We will see a rapidly aging sector of senior developers moving away from the workforce, and no one properly trained taking their places.

Security

The speed of AI code generation introduces novel attack surfaces at scale. Models trained on public repositories can suggest code with known vulnerabilities, like SQL injection or hardcoded credentials; without rigorous human review that amplifies risk dramatically.

Beyond code quality, companies face prompt injection, where malicious inputs subvert the model, and supply chain poisoning, where manipulated packages or training data introduce backdoors. The assumption that "AI will catch its own mistakes" is dangerous current models lack real understanding and cannot guarantee a security posture.

Easier generation of enormous codebases also grows the attack surface by orders of magnitude. Every function, even a correct one, adds an entry point to monitor, and AI-augmented scanners create a recursive dependency: who audits the auditor?

The answer, for now, is to double down on fundamentals least privilege, input validation, dependency pinning rather than expecting automation to solve it.

Infrastructure

Infrastructure is the main problem at the moment. LLMs, with all the improvements, are still a very hardware demanding technology, so much so that the whole market is stuck in a perpetual shortage of many consumer electronic parts, the most prominent being GPUs and RAM memory.

The bet of a whole section of the global economy relies on building more data centers to feed the ever increasing hunger for resources that the multi-trillion parameter frontier models require.

Which brings problems of its own, like a circular money injection stunt being pulled by the major American frontier labs and Nvidia, which might lead to yet another monumental bubble explosion. There is pretty good evidence about that⁶, and it could be as catastrophic as, if not more than, the 2008 economic disaster.

The Future of Software Development

For most engineers, the AI era will not be defined by dramatic productivity gains or the lack of them. Strip away the 10x marketing and the doom-posting, and the honest summary is uncomfortable: the technology is real, the shift is permanent, and the outcome is still undetermined. If you do not sell chips or model access, your bottom line has barely moved. But the ground beneath the profession has.

What actually changed is this: software now sits on a spectrum of value, and the industry has finally learned to see it. Before AI, a dashboard for the sales team and a core system at the heart of the company demanded the same caliber of engineer. Today, the sales team can build that dashboard themselves. That is fine. Not every piece of software deserves the same investment, the same review, or the same future-proofing. Recognizing that distinction is not surrender to AI. It is the first honest act of the new era.

That does not mean quality is obsolete. It means quality is now selective. The frenzy from upper management, drooling over 10x promises, has led the market to forget basic tenets: more code is not better code, and more lines mean more maintenance and a wider attack surface. AI produces more lines faster than ever. It does not produce fewer of them.

So what is left for the developer is the part that got harder: understanding ever-broader scope, deciding what deserves to exist, and owning the result. The machines can write the code, but they cannot be responsible for it. When a dashboard holds up fine while the core system fails at 3 AM, the call still goes to a human. The value of an engineer now lies in judgment. This means knowing when to let AI generate scaffolding and when to intervene with intentional, handcrafted logic. The future belongs to those who can best navigate this new spectrum of abstraction. The AI era has redefined the job away from typing and toward thinking.

References