Vibe Coding at Scale? Engineering Strikes Back - Unladen swallow - Olivier Wulveryck

Skip to content

owulveryck's blog

Vibe Coding at Scale? Engineering Strikes Back

Olivier Wulveryck — 2026-06-19 — https://blog.owulveryck.info/2026/06/19/vibe-coding-at-scale-engineering-strikes-back.html

Generative AI has transformed how code is produced. In just a few months, we went from autocomplete to agents capable of writing, testing, and deploying entire applications. The market is now flooded with methods for framing these agents and making them produce quality code.

But this abundance raises a question few organizations are asking yet: what happens when you are not building one app, but fifty?

The answer lies in a distinction everyone underestimates. A piece of software can be well-crafted without meeting the expected standard. Structured AI methods can achieve a state of the art — the generic one the industry agrees on. But an organization’s state of the art is its own: contextual, agreed upon, living. And that is exactly the one AI does not know — and reinvents, poorly, on every project.

“State of the art” ≠ “YOUR state of the art”

Let us start by clearing up a fundamental misconception: there is no absolute standard of quality .

As Christophe Thibault reminds us in “Done with ‘Technical Debt’”, “Quality” as a singular, universal concept does not exist. At the Plaza Athénée restaurant, you get an excellent meal for around €380. At McDonald’s, where lunch costs under €8, there is also a quality department. Two radically different standards — and each perfectly legitimate in its context. Swap their criteria, and you get two absurdities nobody would pay for.

A state of the art is the set of practices a group agrees upon — explicitly or tacitly — to produce the best possible solution in its context, with its constraints. It is neither universal nor fixed. It is debated, tested, and renewed.

But structured AI methods carry their own state of the art: the one distilled from all the generic know-how accumulated across the internet, libraries, and frameworks. That state of the art is valuable. But it is not yours. And therein lies the challenge: an application can be flawless by generic standards and yet completely off the mark by your organization’s rules.

This is why we distinguish three approaches to AI-assisted development — not by the tools they use, but by the state of the art they guarantee .

Vibe Coding is fast. You prompt, iterate, and accept whatever “looks like it works.” It is the ideal approach for a prototype, a hack, an exploration. But quality depends entirely on the prompt and the developer. No guaranteed state of the art, no reproducibility.

Structured AI assistance (of which BMAD is a good example) goes further. It enforces templates, rules, and detailed prompts. The result is a well-built application, coded “by the book.” This is what we call contextual certainty : for this project, in this context, the result is reliable.

Agentic engineering introduces systemic certainty . Reliability is no longer carried by an individual or a prompt — it is carried by a platform that makes the organization’s state of the art available to every agent, every project, every time.

The differentiator is not whether you use AI: it is which state of the art your approach guarantees. Inspired by Figure 3 from 1.

What structured AI actually does — and what it does not

Let us be fair: structured AI is a genuine step forward. It is not the enemy. It brings a rigor that vibe coding cannot offer, and produces well-made applications.

But it has two structural blind spots.

First blind spot: it imposes its own build chain. Its templates, conventions, and steps are designed outside the organization. They ignore existing processes — the very ones that, in a large organization, guarantee the overall quality clients expect. Those processes often need updating, certainly. But they cannot be bypassed.

Second blind spot: the certainty it produces is local. It says nothing about the next project. The next developer, the next agent, the next project starts from zero. Best practices may be followed, but not necessarily the organization’s rules. The code is well-made, but not necessarily consistent with the other applications in the same system.

This is where a dynamic Jerry Weinberg described long ago comes into play:

The First Law of Technology Transfer: Long-range good tends to be sacrificed to short-range good.<br>— Jerry Weinberg, Quality Software Management

Structured AI optimizes the short-term good (this application, delivered fast and clean) at the expense of the long-term good: system-wide consistency. For one project, this is invisible. For fifty, it is devastating.

The real danger at scale: reinventing the commodity

Here is the precise mechanics of chaos.

When each agent builds its...