Building Products with AI Agents
March 7, 2026
TL;DR: Using an old, abandoned idea, I tried to understand how Spec-Driven Development (SDD) allows to bypass the limitations that typically prevent ideas from being realized when operating at small scale. I have seen firsthand how AI makes it possible to shift focus from code to functional reasoning.
A few years ago, my co-founder and I identified a significant product opportunity. We went as far as evaluating the purchase of a competing product to bridge the gap, but the deal didn't materialize. Ultimately, we shelved the idea because we lacked time and budget required to move from a validated concept to a functional MVP. Recently, I decided to use parts of that concept as a greenfield sandbox. My goal was to experiment with a modern development workflow in a known environment to see if current AI tools could remove the resource constraints that originally made the project unfeasible.
My initial use of AI tools was tactical—refactoring legacy code or hunting bugs. While helpful, this didn't fundamentally change the product development lifecycle. A breakthrough came when I shifted the focus from code generation to requirement clarification. I found that AI is a great tool for reasoning about what a software needs to do and how, at the feature level. Instead of prematurely worrying about how to make the code, the AI helps provide structure to nebulous ideas through iterative Q&A sessions. This allows for a top-down refinement process where complex domain logic is organized before any technical work begins, keeping the focus on clarifying the what before deciding the how.
To perform this experiment, I used a two-stage process: first, the product specification, then implementation using spec-kit (extended with prototyping if necessary). Before touching any tools, I started creating a product specification. This is a comprehensive Markdown document (distinct from any technical files) that defines the high-level vision and domain logic. Throughout the lifecycle, I constantly align this document with the evolving project. Once the product specification is established, I leveraged GitHub’s spec-kit to adopt a formal Spec-Driven Development (SDD) approach to develop features incrementally and iteratively. This is a workflow that starts with a constitution step followed by multiple iterations of sequential steps: single feature specification, definition of the implementation plan, generation of a list of ordered tasks, and execution of the list of tasks. There are other steps available to help clarify edge cases in the spec or to analyze the consistency of the documents created during the iterations.
In this process, the specification is the primary engine of progress. The process begins with a comprehensive Markdown document acting as the "Source of Truth". It forces problem-solving within the knowledge domain first, preventing premature technical choices. This spec provides the framing for every subsequent feature, ensuring coding task are always aligned with broader product goals. Once the specification is robust, coding becomes the final and most mechanical step. When an agent builds from a well-defined, human-steered spec, the output is significantly more precise.
While SDD provides a powerful structure, it hits strictly technological challenges where the best path isn't clear from a domain perspective alone. In these instances, SDD must be coupled with rapid prototyping to explore the solution space. AI serves as a "reinforced search engine" here, rapidly investigating alternatives and generating proof of concepts to validate implementation strategies. When a prototype reveals a new constraint, the product specification is updated to reflect that. This creates a feedback loop between the implementation and the specification, allowing corrective action to be taken in the case of inconsistencies.
This experiment reveals that AI agents are highly effective tools to integrate into existing project management workflows before the coding steps begin. The output of a spec-kit workflow bears a clear resemblance to a WBS (Work Breakdown Structure) produced by traditional project management. By using agents to help with work planning and task handling, we can minimize the number of people required during early stages and we can minimize the need for pure developers, allowing engineers to evolve into architects or management roles.
AI agents are most effective when supporting a technical lead with product management sensibilities. To truly unlock their value, the "human in the loop" must possess the domain expertise to guide the AI, the management skills to coordinate the various steps, and the architectural oversight to keep specifications and code in sync. When these skills are present, AI agents become a powerful force multiplier. They don't replace high-level thinking, they reward it. By treating coding as the final mechanical step, we can focus our energy where it matters most: on the product vision and the user's needs.