From Code to Spec

By Michiel Kouwenhoven, senior developer

We have moved past the limitations of the traditional monolithic approach. In such systems, all functionalities are built as a single complex, interwoven whole that is often rigid and difficult to adjust. Instead, we are making way for a powerful, composable approach. This involves breaking software down into independent, transparent building blocks. The result? An optimized foundation of reusable components and smart 'accelerators' that allow us to easily integrate the best-in-class solutions from the market.

While this is a significant step forward, we are still working with a static codebase. Over time, dependencies age and integrations change, causing even the most modern architecture to slowly turn into 'legacy' if no maintenance occurs.

From Code to Spec

In the traditional world, both the code and the documentation must be maintained. The problem? Today’s code is tomorrow’s legacy, meaning a static project is inherently lagging behind.

In the world of AI, 'spec-driven development' is one of the emerging terms. This means a developer first writes a specification, based on which AI then generates the code.

A specification consists of one or more structured documents, written in human language, describing the purpose, functionality, and other requirements of the software. It is essentially the instruction set that the AI agent uses to start programming. Depending on the level of implementation, the code is no longer the source of truth—the specification is. Maintenance shifts to keeping the specification up to date.

The Agentic Workflow

An AI agent is not a simple chatbot but an autonomous system capable of performing complex tasks through planning, reasoning, tools, and actions.

The agent understands the specifications and, during the planning phase, creates a structured plan in which it divides tasks into several concrete, executable steps. For each step, the necessary information is gathered from various sources. If deemed necessary, provided tools (such as a web search or a call to an external system) or information from a previous step can be utilized. Crucially, there is no need to reinvent the wheel: existing code and previous lessons can be seamlessly integrated.

Only after this plan has been established does the agent begin execution. It is a process where the end result is step-by-step expanded and improved as the process progresses.

Autonomy with Guardrails

When people think of AI, they often immediately think of 'hallucinations' or unpredictable code. In this setup, however, we opt for low autonomy for the agent. This might sound restrictive, but it is exactly the strength of this method: the agent follows exact instructions and works within strict frameworks. Instead of guessing, the agent knows what to build, how to build it, and in what order.

Through this human-in-the-loop approach, the user always remains in control. You get the flexibility of user input—such as preferences for accelerators or specific styling—but with the predictability of the organizational context. The result is a starting point that is immediately client-specific. Whether you choose Magento, Storyblok, or a search engine like Algolia, the agent integrates the right accelerators based on your unique stack.

Business Value: Faster, Agnostic, and Future-Proof

The shift from 'pre-built' to 'generating' delivers direct gains for the business:

• Shorter Time-to-Market: You start on day one with the features that add value for the end user: your unique stack of commerce platform, CMS, and search, utilizing the design tokens from your global design system.

  A global design system safeguards your brand identity through design tokens: the smallest building blocks, such as color and typography. These standards guarantee a consistent brand experience.

• Reduced Technical Debt: Because the foundation is generated using the latest specification, you don't start with a backlog. Maintenance shifts from manually updating outdated dependencies to keeping the specification current.

• Technology Agnostic: Technical requirements are described in the specification. Therefore, a change in technology only requires an adjustment to the specification—the source. This approach makes the setup technology-independent, allowing you to remain agile and respond very quickly to market changes.

Conclusion: The Future is Agentic

This is only the beginning; the true power of this approach lies in its scalability. Think of integrating security checks (OWASP), generating initial content, or writing entire test suites. Once this foundation is solid, you can loosen the reins and lean more on the autonomous power of agentic AI.

In the rapidly changing field of AI, it is uncertain exactly what the future holds, but whether the bubble bursts or not, AI is not going away. With the arrival of agents, a shift is occurring from a chatbot generating content to a digital employee capable of performing effective work—whether you start your next project by writing a specification or not.

At Kega, as a digital agency, we are not standing still; we are exploring these developments and are working to bring them into practice.