The AI Design Authority
We are at a turning point in software development. The discussion often revolves around which AI writing the best code (Claude vs. ChatGPT) or where where AI should reside (IDE or CLI). But that is not the right question.
The problem is not the generate of code. It is the validation of it.
If we embrace AI as “Vibe Coders” – where we specify the intent and the AI does the execution – we create a massive flow of new software. A swarm of AI agents can generate more code in one minute than a senior developer can review in a week. Humans have become the bottleneck.
The solution is not more people. The solution is a AI Design Authority.
From Craftsman to Factory Director
Traditionally, the “Design Authority” is a small group of architects who meet once a week or month to approve or reject a design. In a world of high-velocity AI development that model is hopelessly outdated. It is too slow and too reactive.
If we switch to “Disposable Code” – software that we do not endlessly refactor, but discard and regenerate when requirements change – then our role changes fundamentally. We are no longer masons laying stone by stone. We are the architects of the factory that prints the walls.
But who checks that those walls are straight?
The “Gauntlet”: An Automated Fire Test
An AI Design Authority is not a person, but a pipeline. A “Gauntlet” where every line of generated code must fight its way through to reach production. This process does not replace the human code review with nothing, but with something better.
It works in three layers:
1. The Executive Power (The Generation)
We don't ask one AI for a solution, we ask three. We let Gemini 3, GPT-5 and an open‑source model (such as Llama) work in parallel on the same problem. This prevents tunnel vision and breaks the “laziness” that LLMs sometimes suffer from. This approach is also scientifically researched and demonstrates that you can prevent AI hallucination and build very long chains without errors
2. The Hard Filter (The Law)
No discussion is possible here. Code must compile. Linters must not complain. And crucial: the Black Box Tests must succeed. We do not test whether the function works internally (that could manipulate the AI); we test whether the system does what it should do from the outside. Does the test fail? Straight to the trash.
3. De Zachte Filter (De AI Jury)
This is the real innovation. The remaining solutions are presented to a specialized "Voting AI". This agent does not write code, but leest code. It is trained on our architecture principles, security requirements (OWASP, ISO) and compliance rules (EU AI Act).
Hij stemt: “Oplossing A is sneller, maar Oplossing B is veiliger en volgt onze microservices-architectuur beter.”
De winnaar gaat naar productie.
The Separation of Powers of Software
This model enforces a separation of powers that is missing in many teams.
- The Legislative Power (The Architect): The Architect writes the “Constitution”. The prompts, the architecture documents (
project-description.md, rules.md, skills.md en principles.md), the hard requirements. The architect determines what we build, who builds it, how and why. - The Executive Power (The Coding Agents): They execute. Fast, cheap and under the auspices of human developers.
- The Judicial Power (The Design Authority): An independent AI layer that complies with the law.
Conclusion: The New Role of the Architect
It frees us from the tyranny of syntax errors and lets us focus on what we are good at: systems thinking, truth-finding, structure and decision-making.
The question is not whether AI can write our code. That issue is already settled. Code is largely becoming a disposable product.
The question is: Do you dare to let go of control over the code to release, thereby gaining control over the quality to win back?
let me know
