Development
Three worlds of modern software development: what is really changing and what (for now) stays the same
Most articles about AI development focus on the distant future. Instead, I'd like to take a human look at what's happening today – on real projects. After nine years at the intersection of development and product design, I feel that a sober perspective is now more needed than an enthusiastic one.
In our world, development today is divided into three worlds: vibecoding, agentic development, and classic development. Let's start with the most accessible one – and also the most frequently misunderstood.
Vibecoding
This state was first named by OpenAI co-founder Andrej Karpathy in his tweet from February 2025.
“There's a new kind of coding I call 'vibe coding', where you fully give in to the vibes, embrace exponentials, and forget that the code even exists.”
The key part is “forget that the code even exists.” In practice, you throw a rough brief into a tool like Lovable or Figma Make – and within minutes you have a working app. Without writing code, without knowing syntax.
The result often looks surprisingly professional. But as soon as you want something truly specific, frustration can set in.
AI turns out to be surprisingly weak at fixing bugs. It often gets stuck in a loop – fixing the second bug brings back the first. You can spend a lot of time debugging without a satisfying result.
Another weakness is design, which AI still doesn't handle well – especially subsequent adjustments. If you have a clear vision of the result, it's hard to achieve it.
In short: for perfectionists, vibecoding can be somewhat painful. But if you're willing to lower your expectations and adapt to what AI proposes, it can be a very powerful tool.
Where it works best
Idea visualization and prototyping
An idea you need to quickly show a client, or compare two approaches to the same problem. A conversation over something clickable is significantly more effective and leads to better results.
Micro-projects and websites
A simple dashboard, landing page, client form. Things that used to wait months for development capacity now get done by Friday. Just don't have overly high expectations.
Quick idea validation
Frameworks like Design Sprint or Lean Startup have received a serious dose of steroids. The cost of creating a functional prototype that looks like a polished product and can be tested with real users has dropped significantly.
Where it falls short
Poor foundations
Generated code is great for a prototype you test with potential clients. But as a foundation for more complex products, the quality isn't sufficient. If you decide to keep building on a vibecoded project, expect to rewrite it anyway. The longer you delay, the more expensive it will be.
Security and data protection
Code quality is especially apparent in security and stability. If the application handles data that matters to you, or your business depends on it, vibecoding is not the right choice.
Agentic Development
While vibecoding has been with us for some time, agentic development is a relatively new term. In vibecoding you leave code quality responsibility to AI; in agentic development, the developer still holds it. They orchestrate one or more “agents” – AIs that create something specific. It's like having a team of less experienced developers: splitting larger tasks into smaller ones, designing architecture, and assigning appropriately challenging work whose outputs they review.
According to my optimistic estimate, we're somewhere around 15–25% time savings. But don't tell the developers I wrote that – they're not ready to comment on those numbers yet.
Speed vs. quality
The yolo approach – agents write, the person barely reviews and immediately approves. Time savings on paper are huge, but the result approaches vibecoding with all its downsides.
The perfectionist – the developer reads every line and fully understands it. The resulting code is solid, but the savings can be zero – sometimes slower than writing code by hand.
Classic Development
There are still projects where the cost of failure is unacceptable: eGovernment, banking, healthcare, payment systems, energy, core products of large companies. Here, security, stability, and long-term maintainability are priorities – and that hasn't changed.
What this means
Thanks to agentic development, custom software is becoming accessible to companies that wouldn't have considered it before. Building an internal system precisely for your processes now makes economic sense. That's a real opportunity worth taking seriously – but still leave it to someone experienced.
Vibecoding is here for you too. Smaller projects that aren't critically important can now be done with basic technical knowledge. And for validating or visualizing an idea, AI is an excellent tool.
At the same time, be careful: if someone promises you a critical internal application built with AI at an attractive price, I'd be very cautious. It's likely vibecoding rather than real development. Code that nobody understands and nobody can maintain is a ticking time bomb.
Written by
Jakub Gräf
Published
21. 4. 2026
Written by
Jakub Gräf
Published
21. 4. 2026
