AI coding still needs an architect
AI coding takes the first fast option and stacks workarounds when reality pushes back. Why architects are still part of the equation.
AI coding is great. I mean it. I’m thrilled. You can build serious projects fast, efficiently, without a team. Development can go 100 times faster.
But there are limits. For now, architects are still part of the equation.

What I see
AI looks for the optimal way to solve the task. That’s logical. It takes the first option that looks fast and solves the task as stated.
Where it breaks
Having taken the first option, it starts working along it and goes deeper and deeper. The problems start when the solution leaves lab conditions. In real life, that’s the typical case.
Take enterprise. The ecosystem comes with its own constraints:
- VPN and Zscaler
- security policies
- a limited set of allowed software
- restricted network access
The house of cards
So the AI picked its path at the start, walks down it, and then meets reality. Now it starts inventing ways around the constraints it ran into. That’s how the house of cards gets built. One workaround. Then a second. Then a third. In the end the whole thing is sticks and duct tape.
AI doesn’t see when it’s time to step back and change the architecture pattern because the chosen strategy isn’t working. To the AI, stepping back looks like a change of concept and a huge amount of work. Of course it does: after that many workarounds the thing is already a Frankenstein’s monster. So it stubbornly keeps going down the chosen path.
Where a human wins
There are still limits: the size of the context, and the view of the solution as a whole, the strategic one. This is where a human wins, on accumulated experience. For now.
Discussion
What do you think? How often do you hit Claude's limits?