This article is also available as a video on the Modern Software Engineering channel
Have you had your “Deep Blue” moment yet? That shocking insight that the AI tool that appeared on your desktop some time ago has now got so good, it’s actually become better than you at software development. Just like Garry Kasparov when he was beaten at chess by IBM’s Deep Blue. Your carefully cultivated skillset which has brought you success and quite frankly a pretty good salary these many years, seems worthless now a machine is just as good. Agentic Workflows changed EVERYTHING in 2026.
Deep Blue is the name given to this existential dread you can get from the way AI is taking over software development. This term was coined by Simon Willison and others on a recent podcast and blog. Not long ago, the quality of the code these AI tools would come up with made them too unreliable for anything important. That has changed now. More and more developers are realizing these LLMs and agentic tools can do large parts of your job. All that syntax that you’ve memorized, those refactoring shortcuts that flow from your fingers, some of those skills may be obsolete now.
It’s scary, especially if a lot of your self image is bound up with “I am a person who writes code”. According to Steve Yegge and Gene Kim in the foreword to their book “Vibe coding”:
“The good news is you’re not too late … yet”
Talk about FOMO. Of course they are trying to sell their book, and, there is some truth in this. Things are moving really fast.
Some still refuse these tools
There are some developers who I know well and admire, who are still refusing to use these tools, not because they are scared, although they probably are that too, (we all are), but if you ask, they often cite ethical reasons.
This is a legitimate concern, – given the way LLMs are trained on copyrighted works, exploit low paid workers and consume a huge amount of electricity, water and other precious resources. Many large language models are being controlled by American tech billionaires who are not politically neutral.
I myself was slow to take up using AI tools partly for those reasons. As a software engineer in 2026 I’ve accepted that AI coding tools are now part of the job. Much like being a member of modern society means buying new clothes, flying and eating meat. These are also questionable ethical choices. Some people choose to avoid these things completely and I admire their principles, although personally I choose differently. I don’t farm my own vegetables and take the train everywhere. I make better choices where I can. I hope we all do.
Sometime in the past few months the AI agents got good enough that most software developers will do a better job if they learn to use them. Everyone in this business more or less has to face up to that Deep Blue moment, and make a choice. Agentic workflows have changed the job.
Let’s take the positive view then, what are the skills you already have as a senior developer that transfer well if you adopt agentic AI? Well, it could affect the architecture.
Winchester Mystery House
We used to have two approaches to architecture – the Cathedral and the Bazaar. It’s an analogy by Eric S Raymond from 1998
- The Cathedral model is carefully planned, closed-source, and managed by an exclusive team of developers.
- The Bazaar model is open, transparent, and community-driven.
The Cathedral model was the dominant approach until the internet made the Bazaar model possible. Today a third model has emerged, enabled by AI. Drew Breunig pointed this out: the Winchester Mystery House architecture model.
For those of you who are unfamiliar, this is an actual building in California, now a famous tourist attraction. It was designed and built over a long period by someone with a keen interest in architecture and interior design, and enough money to make any and all their ideas reality. It’s all been built and rebuilt many times, all the rooms, all the kitchens, all the bathrooms, all the staircases – using the finest materials and craftsmanship, and a huge amount of flair and innovation. Newer parts made older parts obsolete without removing them. Staircases that end in a ceiling. Windows that became barred up. Emergency “quick fixes” after earthquake damage were put in and never revisited. Sound familiar? It’s called a mystery house because it is so confusing, impractical, charming and idiosyncratic.
With today’s AI tools it’s easy to come up with something like that.
Modern Software Engineering Essentials
Again I come back to the essentials of modern software engineering: managing complexity and optimizing for learning. I see a good future for those skills and expertise.
As developers we already know how to manage complexity through finding abstractions, making re-usable pieces and keeping the codebase organized and readable. Agentic AI doesn’t just do those things spontaneously. Birgitta Böckeler has written a great piece about Harness Engineering. We need to create ‘guides’ and ‘sensors’ that are specific for the code we are working on, and our team and our organization.
When it comes to optimizing for learning – this is where we still need to understand what’s needed, listen to users, and build in feedback loops so we can adapt as we learn more about the problem we’re trying to solve. Behaviour Driven development is a key skillset here.
Skills we no longer need?
It’s harder to identify the skills we no longer need. In the past, software developers needed to know how to make punch cards and manually patch them with pieces of tape. Nobody does that any more, except in museums. There are probably skills like that now. I hesitate to pinpoint them though. It feels like it’s too soon, and many of our skills are learned together and depend on each other.
For example, people have pointed out we’re not writing code by hand any more so perhaps we don’t need to learn that? Although at the moment, we still often read the code, and these two skills – reading and writing sourcecode – are hard to separate.
Don’t need to read code any more?
I heard an intriguing anecdote about a new graduate developer with very little coding experience who became very rapidly productive using agentic AI. He was shipping lots of features and being a useful member of a small development team. Even after some months though, apparently he still couldn’t really read sourcecode very well.
How could that work? The crucial success factors seem to be firstly having a strong Agentic and project specific Harness already in place, with good guides and sensors including tests. So he could generally rely on the agent to produce good code for him. Also, he did a lot of pairing and observing other experienced developers working with these tools. He had strong mentoring in engineering skills. It worked in that context, to largely drop code reading skills.
I don’t want to advise you not to learn to read and write sourcecode, I fear a future of burgeoning Winchester Mystery House architectures. It’s pretty early days for agentic tooling and some problems will likely not surface for some time.
This graph does give me pause though. It shows the time a human would need to complete a task, against the date at which a LLM was able to complete that task. They are getting better and better at doing larger tasks unattended. Notice it’s a log scale, showing an exponential improvement that shows no sign of slowing down. We humans are very bad at intuitively grasping exponential change. In a few months or years, maybe reading code will become a specialist skill for museums. Exciting times.
Conclusions
What I want to leave you with is some optimism. Yes, the job changed. Senior Developers are not dead yet though. You have lots of transferrable engineering skills needed in agentic workflows. Harness Engineering for example. Face your Deep Blue moment and come out the other side.
Happy Coding! (with agents!)
