The Architecture Gap: What AI Still Cannot Do
Eight years thinking. Three months building.
That is the story of Lalit Maganti and syntaqlite—a SQLite parser, formatter, and verifier that they had procrastinated on for nearly a decade. What changed? Claude Code.
But the story is not what you think.
The Prototype Was Easy
Maganti had a problem: 400+ grammar rules to work through. Pure tedium. The kind of work that makes you put things off indefinitely.
AI obliterated that barrier. The first prototype came together fast. A proof of concept that worked. Something to play with, break, iterate on.
AI basically let me put aside all my doubts on technical calls, my uncertainty of building the right thing and my reluctance to get started by giving me very concrete problems to work on.
This is the promise of agentic engineering made real. Not replacing thought, but eliminating the friction that prevents thought from becoming action.
Then Came The Hard Part
The prototype worked. So they shipped it.
Except—they did not. They threw it away and started over.
Why? Because AI had produced code, but not coherence. The low-level pieces worked, but the high-level architecture was a mess. Decisions kept getting deferred because "refactoring was cheap." The codebase stayed confusing while the design corroded.
I found that AI made me procrastinate on key design decisions. Because refactoring was cheap, I could always say "I will deal with this later." And because AI could refactor at the same industrial scale it generated code, the cost of deferring felt low.
This is the architecture gap. AI can implement anything you describe. It cannot decide what you should build.
Implementation Has Answers. Design Does Not.
The key insight:
Implementation has a right answer, at least at a local level: the code compiles, the tests pass, the output matches what you asked for. Design does not. We are still arguing about OOP decades after it first took off.
When the problem has an objectively checkable answer, AI excels. When it does not—when success is a matter of taste, tradeoffs, and long-term thinking—AI wanders. It follows you down dead ends you did not know you were taking.
When I was working on something where I did not even know what I wanted, AI was somewhere between unhelpful and harmful.
The Pattern
This keeps showing up:
- AI crushes well-defined implementation work
- AI struggles with architectural judgment
- The more uncertain the problem, the less helpful the agent
The solution Maganti found: use AI for what it is good at, but own the design. The second version of syntaqlite took longer. Required more human-in-the-loop decision making. But it produced something robust.
What This Means
The coding agent hype cycle talks about replacing developers. This story points elsewhere.
AI does not replace architectural judgment. It amplifies the consequences of having it—or not having it. If you know what you want, you move faster than ever. If you do not, you just get lost faster.
The skill gap is not prompting. It is knowing what to prompt for.
The real productivity unlock is not vibe-coding your way to a prototype. It is having the clarity to know which prototypes deserve to exist—and the discipline to throw away the ones that do not.