The Architecture of Agency: Moving Beyond Vibe Coding

Lead Architect Carlos Kelkboom challenges the industry trend of vibe coding, where engineers rely on AI to generate code in isolation without a clear structural plan. Kelkboom argues that while AI agents handle one-shot tasks well, they fail at iterative engineering because they cannot maintain a coherent system architecture. The hidden risk of current AI workflows is a shift toward individual isolation, where engineers lose the ability to collaborate, mentor, and internalize technical knowledge. Software engineers and technical leads should adopt Kelkboom’s scaffolding-first approach to reclaim agency, turn AI into a force multiplier, and build software that is deterministic, maintainable, and designed to last.

The Hidden Cost of Fast Solutions

Most teams treat AI agents as code generators, asking them to build features piece by piece. Kelkboom identifies this as a fundamental error. Agents are poor at iterative changes because they lack a long-term memory of system constraints, often duplicating logic or creating God objects to satisfy immediate prompts.

The immediate benefit of this brute force approach is speed. The downstream effect is a codebase that becomes unmaintainable within months. By skipping the design phase, engineers trade long-term durability for a short-term gain.

> "They are good at writing code but not at engineering and application."
>
> -- Carlos Kelkboom

The 18-Month Payoff: Why Scaffolding Wins

Kelkboom’s methodology requires 20 to 30 minutes of manual scaffolding, such as defining folder structures, function signatures, and data boundaries, before an agent touches a file. This is the unpopular but durable path. While most developers want to jump straight into prompt engineering, Kelkboom insists that the architecture must exist in the human mind first.

This creates a lasting advantage: because the structure is defined, the agent is constrained to work within the system rather than creating the system. Over time, this prevents the hallucination of architecture where an agent decides to replace your deployment scripts or business logic without warning.

> "If I don't understand it, I throw it away. I don't even try to go much deeper than that because I know that if I don't understand it, it has not continued in the scaffold that I gave it and I know that that it will bite me later if I just merge it into my code."
>
> -- Carlos Kelkboom

How the System Routes Around Your Solution

The current AI coding ecosystem creates a feedback loop of isolation. Because agents provide instant answers, the incentive for junior engineers to ask a senior peer for help or to engage in a rigorous pull request review is evaporating.

Kelkboom warns that we are training a generation to use tools without understanding the underlying mechanics. If the AI bubble bursts or token costs skyrocket, the industry will be left with a workforce that knows how to prompt but not how to engineer. The system is currently routing around human collaboration, which is where the most critical technical and architectural knowledge is transferred.

Key Action Items

• Implement Manual Scaffolding (Immediate): Before starting any new feature or project, manually define the directory structure, class signatures, and data flow in Markdown. Do not let the agent create the structure for you.
• Adopt the Open Mode Workflow (Immediate): Force your primary AI agent to never touch files. Use it only for conversation, planning, and teaching. Only spawn sub-agents to perform the actual coding once you have a 90% confidence level in the plan.
• Use Git Worktrees for Parallelism (Next 30 Days): Instead of forcing one agent to iterate on a single branch, spin up three parallel Git worktrees. Let three different agents attempt the implementation, then use a fourth agent to compare and merge the best patterns.
• Prioritize Deterministic Languages (6 to 12 Months): Begin experimenting with languages or subsets of languages that avoid heap allocation and dynamic memory. Kelkboom argues that deterministic, bounded systems are the future of reliable AI-driven development.
• Reintroduce Human Review Cycles (Ongoing): Even if an agent produces the code, force a human review process that mimics traditional PR cycles. If you don't understand the code the agent produced, delete it and start the conversation over.
• Build Local Knowledge Graphs (12 to 18 Months): Move away from centralized databases for your project context. Use a distributed system of Markdown files with indexed front-matter to create a searchable, human-readable map of your project that agents can query.