Vibe coding, six months later

In March, I wrote here that AI should remain a tool serving sound engineering, and I told the story of an internal application generated almost entirely through prompts, brilliant in demonstrations but fragile as soon as the volumes became realistic. My conclusion was cautious.

Six months later, our practice has nevertheless evolved, and from the outside it might look like vibe coding: we describe what we want in natural language, and agents produce a large part of the code. I therefore owe you a follow-up, explaining what has changed and what has not.
 

What an outside observer sees

A developer expresses an intention. One agent makes a proposal, another checks it, and a third tests it. The code appears quickly, sometimes very quickly. If we stop at this image, we might think we are witnessing what I described in March. The result, however, behaves differently.
 

What happens before the first instruction

Every intervention by an agent relies on a framework built before it. An agentic architecture: specialised agents, each with a specific skill. One retrieves the design system elements, another analyses the tickets and proposes an order based on dependencies, another builds the administration screens, and another tests and checks security. The developer remains in control: they provide the intention, review and validate.

This framework draws on twenty years of web projects. Our architectural choices, coding standards, ways of structuring content, managing permissions, preparing a deployment or testing a feature have become instructions that the agents apply. They are cooking recipes in the literal sense: they have been tried, corrected and refined on real projects, for real clients, long before the arrival of AI.

That is the difference from the case I described: the fragile application relied on an accumulation of blocks generated as requests were made. Here, each block is placed within a solid technical architecture before a single line is written.
 

What hasn't changed since March

• Code is only part of the work. Clarifying needs, making decisions, reviewing, deploying: these steps remain. AI alone doesn't change much about that.
• Review remains essential. The faster code arrives, the more reviews matter. A team that doesn't review what an agent produces is building up technical debt at TGV speed, at Très Grande Vitesse.
• Understanding comes first. No one puts into production a component that the team would not be able to explain, evolve or repair five years from now.

What has changed

The effort has shifted. We spend more time upstream, framing, specifying and formalising what ten years of practice had taught us to do implicitly. We also spend more time downstream, checking. The middle, the actual writing, is speeding up. This shift is not unique to our team, either: DORA's research on AI-assisted software development also observes that the time saved on code generation tends to be shifted towards auditing and verification.

It's not magic, and history is repeating itself. I remember a time when clients told me that their son had built a website and a Facebook page in two days. It was true; the website existed. Everything else remained: making it discoverable, adapting it to the people who use it, securing it, evolving it along with the organisation, and entrusting it to someone capable of taking it over. Twenty years in the profession have taught me that the speed of a first version says little about the value of a project.

Today, we hear that projects will cost two or three times less, and I am wary of that kind of promise: at this stage, no one can say, and it reduces everything to a question of price. What I see is more comprehensive work. Accessibility, SEO and GEO, automated tests: we think about them earlier and more often.

And I remain cautious. This approach works because it is based on prior experience. A beginner who entrusts a critical project to agents without any framework will very probably reproduce the March case. Professional experience is becoming a more valuable asset, not a less valuable one.
 

And what about you?

Do you recognise yourself in this journey? Would you like to accelerate your digital transformation and integrate AI into your business processes, while retaining control over what you build? Perhaps you have already tried things that failed to live up to their promises, or inherited an application that no one knows how to evolve anymore.

Since 2006, we have been supporting organisations in their digital transformation and in setting up ecosystems where tools communicate with one another. What we have learned by putting AI to the test in our own organisation, we are happy to share: what works, what took longer than expected, and what we would do in your place.

Write to us in a few lines about your context, and we will get back to you.

Conclusion: the framework matters more than the prompt

Six months later, I stand by what I wrote in March. I would simply phrase it differently.

The problem is not that AI writes a large part of the code. It begins when we delegate the architecture, trade-offs, validation and, ultimately, responsibility for the outcome to it at the same time.

Viewed from a distance, the way we work today looks like vibe coding: an intention expressed in natural language and a lot of code produced by agents. But beneath the surface lie precisely the elements missing from improvised vibe coding: an architecture, standards, context, tests, reviews and a team capable of understanding what has been built.

The more AI accelerates execution, the more important the engineering that frames it becomes.

Expertise therefore does not disappear. It shifts. Knowing how to produce every line of code yourself may matter a little less; knowing what needs to be built, why, how to verify it and how to evolve it matters even more.

Want to learn more about vibe coding?
 

KEY TAKEAWAYS FROM THIS ARTICLE

• AI can now produce a significant portion of the code, but generation speed guarantees neither the quality, security nor maintainability of a project.
• Agentic development works when it is built on an existing framework: architecture, standards, business rules, tests and experience accumulated on real projects.
• Human responsibility remains central: generated code must be reviewable, understandable, testable, correctable and maintainable by the team.
• AI shifts work more than it makes it disappear: less time spent on pure coding, and more on framing, decision-making and verification.