Prexa blog
The Advent of Vibe Engineering
As per humankind's previous delving into Vibe Coding... we're making the same mistakes with software architecture!
There is a scene in This Is Spinal Tap that has always stuck with me.
The band wants an impressive Stonehenge set for its latest song. One of the band members sketches this idea on a cocktail napkin but he writes the measurements illegibly. The production team dutifully builds a Stonehenge that is ridiculously tiny. During the performance, this petite monument is lowered carefully onto the stage with lights and fanfare but is completely out of proportion to the band playing around it.
What makes it funny is that the mistake is absurd.
But it is also a surprisingly good metaphor for where software architecture seems to be heading in the age of AI. We are becoming increasingly comfortable with what I can call vibe architecture... sketch something rough, provide a prompt or two, and let AI do the rest (ie. fill in the gaps and let'er rip.)
Collectively, we've done this before, in the early days of Agile. We aren't thinking carefully about the moving parts of the system before we start assembling them. Imagine asking a group of architects to design a building in San Francisco. They would think about structure, loads, materials, foundations and ... most importantly, earthquakes. Now imagine inviting a group of carpenters onto the site and providing a cartoonish picture of a building on a cocktail napkin, “Let's get cracking. Time's a wasting. We’ll work it out as we go.”
You're going to get something... but I wouldn't call it a well-thought-out building.
AI makes this temptation worse because it sounds confident and authoritative.
And then we get lazy.
It provides detailed answers. It produces diagrams, code, APIs and deployment plans in seconds.
What's not to like?
Saves us time, and gives us something that we don't have to think much about. Two for the price of one!
Unfortunately we're going to get the Bad and the Ugly along with the Good engineering insight.
Yep, that's because of the way that the AI data is built... it's a proxy for the collective output of humanity. That means it will be referencing the genius of the Golden Gate Bridge right alongside Joe's Substack rant from last Tuesday.
So let's look at the use of AI for architecting in the lens of cost (we will get to cybersecurity in a moment.) It seems that AI has dramatically reduced the apparent cost of writing software. But what if it gives us an architecture that is subtly different from what we are asking for and we aren't providing enough oversight?
We are bound to get endless cycles of rebuilding, refactoring, reintegration and retesting. Now, we -can- perform those cycles much faster. But that's not the point. Every cycle will still consume engineering effort, compute, and especially tokens!
Now imagine that we, being lazy engineers, allow some of the structural mistakes to survive. The cost of removing these issues goes asymptotic pretty quickly as additional dependencies are linked with our architecture.
A problem that takes an hour to correct during design can become a week of engineering work once services, integrations, permissions and operational processes are taken into account.
Unfortunately cybersecurity issues are an accelerant, causing expensive problems to balloon in potential damage.
Poorly considered permissions, trust boundaries, data flows and component interactions eventually become major security headaches. An actual breach may generate enormous downstream costs through incident response, governance, remediation, regulatory exposure and reputational damage.
Oof.
Thankfully the answer doesn't need to be returning to months of heavyweight design before any code is written. It is much simpler... be slightly more deliberate before you start building. Dust off the Pareto Principle and liberally apply it. Spend about 20% of your effort thinking carefully about how the system should operate (business objectives, derived security constraints on those objectives, customer requirements). Then let AI loose on the remaining 80%... implementing the objectives, constraints and requirements.
That small investment in time can easily eliminate a disproportionate amount of problems later.
And you won't have a fear of a 3 foot high Stonehenge replica disrupting your workflow.
