Why you should try AI-assisted software engineering?

For the past few weeks I’ve been working heavily with various coding agents on multiple different tasks – from writing code, fixing bugs but also understanding existing feature implementations as well as specifying requirements and even writing JIRA user stories.

A few observations to share:

Yes, code writing becomes the least of your concerns

“Thank you captain obvious” take – but yes. 15 months ago – when I started to explore that space and tinker an application in the garage – I was still writing 80-90% of the code by hand. It gradually improved over time, but the leap that happened in the last 4 months is insane. Things that weren’t reliable half a year ago simply work now. A POC that used to take 2-3 weeks stitching together pre-built components now takes 3 days and largely from scratch.

Still, you got to know what you want to build.

Making things highly configurable matters less than it used to

If you can change something in minutes with AI assistance, the traditional argument for building elaborate configuration layers starts to look different. Feels crazy to say that but yes – hardcoding stops being a dirty word.

SDLC process evolves

The world we grew up in worked like this – few weeks of discovery / analysis were turned into heavy requirements documents serving as an input to high-level design & scope. These were then decomposed into epics / user stories building up a backlog of to-do items that got prioritised and incrementally delivered passing through development stages.

I don’t believe that ground rules in this philosophy are changing, but the details are. We all need to adapt to the new reality and learn what the new SDLC process will look like.

I don’t have a full recipe yet, but here are a few paradigm shifts that I see:

• Analysis | Requirements Business Analysts become super heros, but they also got some super powers. Spinning 3 versions of a dirty prototype as a way to validate assumptions is now easier than running 2-week discovery exercise.
  
• Architecture | Design Clearly articulating architectural decisions and documenting them in the codebase becomes critical guardrail to keep coding-agents in their lane. Spec-driven development seems to be the way to go.
  
• Testing Unit testing & test automation become very easy. Manual oversight is still required, but more to control whether testing agents are not “grading their own homework”.
  
• DevOps Automation Proper pipelines with security checks and regression tests become more important than ever.
  

The role of Product Manager shifts dramatically

A PM who writes loosely defined requirements and relies on engineering to figure out the details might struggle to find himself in this new reality. I don’t know where it will go, but the gap between ‘defining what to build’ and ‘building it’ is narrowing fast.

I’m not talking vibe-coding

It’s a beautiful term that was coined to describe this new paradigm. But what I’d call vibe-coding is “generate an app for XYZ” – pure prompt, no guardrails, no architectural oversight. It’s very far from what I’m talking about here.

Vibe-coding is asking AI to build something and hoping for the best. AI-assisted software engineering is knowing exactly what you’re building and using AI to get there faster.

You will hear that “it’s not production-ready”

I understand where this comes from, but I think it’s mostly a status quo defence. I have seen genuinely terrible, duct-tape solutions running happily in production for years. With proper oversight and guardrails, the overall quality of AI-assisted code is at least no worse. And often better than what gets shipped under deadline pressure by a tired team.

Yes. Expect to hear stories about security breaches or AI-agent wiping production database. That’s vibe-coding without guardrails, without proper software development lifecycle. Not the same thing.

It’s time to act.

I’m grateful to work in an environment where one of our leaders created a safe space for us to actually try that new approach. I can see some people reacting with resistance or at minimum passive indifference. That’s normal. People tend to stay in roles that feel safe and familiar. But the window to build this muscle is shorter than it looks.

But if you’re a leader at your organisation – it’s time to act now and your job is to push people. Select one crappy legacy app that everyone hates and challenge your team to rebuild it with AI. Most likely it won’t fly, but at least it will make them think how to work in this new reality. If you don’t have any idea how to start – invite a partner that will help you to do that POC hand in hand with your team, so that they could learn.

Curious to hear from others who are doing this seriously – or from those who think I’m completely wrong. What % of code is shipped by AI? What is the future of Product Manager role?

Disclosure: written by a human (myself). AI was involved. Mostly ignored, occasionally adhered to especially when it comes to spell-check and grammar.