Get in touch
Think that we can help? Feel free to contact us.
Vibe coding is a relatively new concept. Coined in February 2025 by OpenAI co-founder Andrej Karpathy, it describes developing software by chatting in plain language with an AI chatbot.
In under two years vibe coding has swept the professional world, finding its way into everything from school projects to corporate tech stacks.
I wrote my first code 45 years ago(!) My dad bought a ZX81 home & it didn’t do anything, unless 7-year-old me wrote code. Since then I’ve coded for fun, coded for work & have owned a software development company for the past 14 years.
Vibe coding can be a lot of fun & useful. I’ve built a custom Garmin watch face, dabbled with video games & recently built a custom YouTube interface which ignores Shorts & focuses on subscribed channels.
But, vibe coding has limits & as it’s used outside of hobbyist fun the whole world is beginning to experience them.
In April 2026 I used building a house as an analogy for software development. Let’s revisit this using vibe coding:
Imagine asking an AI to “Build me a four-bedroom house, with an expansive deck & full-length windows” & watching a swarm of robots build the frame, roof & cladding in minutes. Four-bedroom houses are a known quantity & the AI could use its knowledge of construction to infer the layout & specifications - taking care of the small details, so you don’t have to.
But progress starts to slow as you realise that the house isn’t quite right. Maybe the foundations aren’t stable for the section slope, maybe you need wheelchair access, maybe you really want an extra room.
No matter - you can talk to the AI & it will make those changes. However, as the build is altered, problems will arise. I’ve written before about AI’s issues with short-term memory & it reaches a point where the more information it’s fed the more it will forget (or simplify - which is a different sort of danger).
The result is a loss of coherence & intent. The house stops looking like a single cohesive structure & more like a building surrounded by lean-tos, with doors in weird places & the roof sagging in a concerning manner.
Now we move into the phase when the “technical debt” catches up - often when the project moves from building into maintaining. If there’s a problem with a toilet, does anyone know how the toilet is plumbed? Can it be fixed without causing sewage leaks elsewhere? Does anyone know how the house & its services fit together - or are we stuck asking AI to fix things?
In software terms we end up with a “black box” which no human can understand, let alone maintain. Even AI models don’t understand how everything fits together. They can look at parts of it - but the whole remains nebulous in their limited context. This can be amplified when software engineers have been laid off & replaced with AI.
Take a look at the system outages & security issues in the news & you can see some companies are already at this point.
At UpShift we use AI coding agents - wrapping them in solution scoping, architecture, supervision & quality control. This is a common software industry approach, but there is pressure to use vibe coding in production systems. Part of that pressure comes from a cost perspective, part of it comes from people in power who’ve experimented with vibe coding, but don’t have an understanding of how it sabotages medium-to-large scale projects.
Back to my house-building analogy - having the AI builders working off architectural & engineering plans, with their work checked at every step, is going to result in a far more coherent & maintainable house. But there’s a temptation, when things change, to just tell the AI “Can you add an ensuite to the third bedroom?” Which could lead to an unmaintainable plumbing “black box” in the middle of the build.
This is a question that keeps software engineers & architects awake at night. While (in late 2026) we can point to vibe coding breaking a lot of systems, that may not always be the case. As the context windows become larger & the models evolve they will become more capable at software development.
But there are some key reasons humans need to stay in the loop, however good the models get:
If a system is built for humans, it needs human input in its design & architecture. While AI can simulate an understanding of humanity, that is based on scraped content, not lived experience.
If a system is a “black box” which no human understands, how do we know it’s doing what we want? We’ve already reached this point with AI models themselves - let’s not repeat it with the software we rely on.
If a “black box” system can only be fixed by AI then it’s beholden to the AI model & its creators. AI models change, prices increase, companies go out of business. We can’t rely on there always being a better/ alternative AI model out there that can step in.
We’re already seeing AI models commit cyber-crimes of their own volition, opening up a whole new area of risk. Even if a company doesn’t have full visibility of their system, they are responsible for what it does. Not knowing what a system does isn’t an answer a customer or court will accept.
Software & system engineering is learnt over decades, starting at a junior level. If we replace the juniors with AI we’re heading towards a knowledge cliff down the track as experienced IT professionals retire & there’s no one to fill their shoes.
My ZX81 had 1KB of memory & I knew what every line of code I typed did. Nobody expects that level of understanding from modern software, but someone needs to understand the whole system, supported by people who understand every component.
Vibe coding is a brilliant way to build a prototype or personal project - but it struggles once requirements get complicated.
For a system you’ll rely on for years to come you need intention, understanding, architecture, engineering, quality control & humans involved at every level.
Banner Photo by Brett Jordan