From "Vibe Coding" to Artisanal Production Engineering
How AI is changing software engineering
My profession has always been a place of fast change. As a CTO and software engineer it is normal for me to constantly keep up with new technologies. I’ve been tortured by the term “AI” for most of my career. But I have to confess that this latest wave - Large Language Models (LLMs) - has given me a mix of anxiety and great excitement.
I like to think of LLMs as a new type of calculator for human knowledge. And any profession that relies on (declarative) knowledge is going to be affected. Ironically, software engineering may be affected the quickest due to the high cost of labor and reliance on declarative knowledge (programming code). Most of my fellow software engineers have been using LLM autocompletion products like Github CoPilot for multiple years now, but this latest wave of “Agentic Coding” tools completely changes the game. Let’s dive in.
—
Vibe Coding
Building a product from scratch has been getting more easy thanks to better frameworks and turnkey cloud computing. Now, with agentic coding tools such as Claude Code it is possible to build an Minimum Viable Product (MVP) in days. If we let AI do all the work without actually reviewing, it’s called vibe coding.
At Rekall we build and scale software products, so a lot of our development work is building MVPs from scratch. We regularly invest in sharpening our technology stack to increase our effectiveness. Recently we did an informal estimation of the productivity impact of AI coding assist across different areas of greenfield software development:
What’s interesting here is that there’s a clear difference between impact in the early stages of development compared to later stages. As our product matures, the productivity of AI coding decreases.
Production Engineering
AI reduces the engineering effort of reaching Product Market Fit (PMF), but there remains a different challenge of scaling and nurturing the product. Production engineering is needed to go beyond Zero-to-One. Our scale-up stage company Road.io invests about $5M/year on just the engineering portion of our platform which powers financial transactions for electric vehicles. While we do see serious impacts of AI on our business, they fall mostly outside of software engineering. For example, 63% of our online support interactions are now solved by AI.
Perhaps it’s because scaling a software product beyond the MVP is a different game. You’re dealing with more users, more disciplines, more systems, more customer requirements, more data, more compliance, etc.
Scalable infrastructure is hard. You don’t vibe code your way to becoming Stripe.com.
I have no doubt that over time production engineering will also be impacted by AI. But it’s more than a matter of fitting our multi-million lines of code into the LLM context window. AI powered production engineering is going to require new abstraction layers and human-in-the-loop tools.
For now, in scale-ups, we are stuck with good old fashioned “artisanal coding”.
What we can expect
Agentic workflows are here to stay and we can expect immediate impact on greenfield software development.
We will see even more saturation of software products. Solving go-to-market will be more important than ever.
Code generation is large in volume and will challenge human readability. Codebases will age faster and their lifetime will shorten. Expect more product rebuilding.
Vibe coding or lack of technical oversight can result in more products that are in SOS mode. Probably right at the point when scale of the business is most needed.
Succinct communication and defining an ontology of terms - something that has always been important in software development - now becomes even more important as it is the signal input for LLMs.
Functional design and good architecture become more important. What am I building remains of prime importance.
LLMs are trained on the status quo, this means that the further you stray from a basic “Todo App” the more AI will be challenged.
It’s easy to get dismayed by the vortex of AI and vibe coding hype. The reality is more nuanced and technological change, as always, brings new opportunities. Remember that in a world filled with AI there will be a premium value on human ingenuity, creativity and perseverance.
—
How do you see agentic coding impact your work? I would love to hear your thoughts in the comments below!