Is it over?

2026-10-25

Under the headline:

På 88 timer udraderede AI årtusinders videnskab – »Det er en kæmpe bombe, og folk er fuldstændig i chok

in an article in Berlingske today, 24/10-26, mathematician Nikolaj Thams explains how shaken mathematicians around the world are these days.

The reason: About two weeks ago OpenAI published around 700 papers in which AI agents have solved mathematical problems at a previously unseen high level.

It is another "Deep Blue moment," just far more extensive than when a dedicated chess computer first beat a human, back in 1997.

"It's a massive bomb, and people are completely in shock. If we were to transfer that to journalism, it would be like every article containing a Watergate revelation, which would lead to great careers and prizes for those who published them. And they just toss them out almost like a zip file" [Nikolaj Thams, from the article]

And it has happened quickly: A year ago, AI agents were far behind humans. Now any doubt about whether agents are better than humans at solving these tasks has been completely swept off the board.

I'm sitting with much the same experience when it comes to writing code, which has been a large part of my job as a systems developer. There are some differences - and hope!

Systems development is about more than writing code

A year ago I was skeptical about whether coding agents could even write code of the same quality as the most skilled developers. I have packed that skepticism away. In the same way as the mathematicians above, I am no longer in doubt: The agents produce code that is at least on par with the most skilled developers I have met.

If they are guided correctly.

And guiding agents itself has become a lot easier than a year ago. Tools like Claude Code and OpenAI Codex are much better now, and the time when we sent people on courses to learn "prompt engineering" is over.

Looking back, there is just one thing that stands out: Only about half of my time has gone into actually writing the code. The other half has consisted of:

  • Communication with users, customers, or other key people.
  • Clarifying what is actually wanted (not necessarily what the user/customer says they want).
  • Prioritising and "packing" tasks - what can we deliver in the next sprint, for example.
  • Assessing and choosing solution options.
  • Design and architecture. Devising and sketching proposals that are shared in meetings or otherwise with the user/task giver/system owner.
  • Monitoring running solutions, debugging errors, communicating about these.
  • Writing and maintaining the necessary documentation.

There are some of these points where AI agents can also assist - for example, producing and keeping the right documentation up to date.

But I think we will keep primarily using human intelligence on the points above.

A modern AI agent can generate output almost endlessly, it never gets tired, can deliver 10 different solutions to any challenge - but it is still us who have to assess, analyse, prioritise and communicate with the business, users or customers about the solutions.

We still have to build solutions and we still have to make workflows smarter. We have to choose the right things to build.

Just as we can get stuck in the supermarket if we get too many options - "paralysis by analysis" - and end up not buying a jar of jam because we were faced with a choice of 13 different flavours, we will have to become good at choosing exactly the right combinations of components in a solution.

So we are not quite as challenged as a theoretical mathematician, I would argue. Because we are also practitioners as systems developers. Something has to be built and that is very different than finding correct answers to complicated theoretical problems.

New possibilities

AI agent-driven development has given me the opportunity to take on things that were not previously possible.

A few examples of solutions I have made this year, that would not have been solved a year ago:

  • We have dropped a large American SaaS product for project/task management in the IT department and built our own solution. The first version was live in 21 hours, and we now have exactly the functionality we need - neither more nor less. Plus, 7 x $21 is saved every month.
  • I have made my own news agent that scans RSS feeds for interesting AI news and sends me an email every morning at 06:00. The first version of this solution was made in 2 hours, lying half-dozing on my sofa, under a blanket, 100% from my iPhone (/remote-control in Claude Code). Including deployment of an Azure Function and Table Storage, integration of Resend for sending the email, and calls to the Haiku 5.5 model for summaries, de-duplication, translation, etc.
  • This blog - and the site behind Biilmann.dk. Something I have long wanted to build but have not had time for. Now I have a blog that uses Tina CMS to manage all posts as markdown files in a GitHub repo, uses Haiku 5.5 for automatic translation (draft) into English, and automated GitHub workflows to release new articles (without me having written the underlying YAML files for this).

These are all examples of things that would have been projects lasting several weeks or months before I had AI agents produce code.

Is it over?

So is it over for writing code?

I will not rule out that there is some code that still has to be written by humans - but right now I can see that I am very quickly moving away from it in my daily work.

But being a systems developer is not over. There is still a huge demand for good solutions - in fact, I would argue that demand has risen, because AI is being hyped enormously in the media, online and in the corridors of management. The promise of enormously increased productivity itself causes expectations and demands to rise.

So I still don't believe we will reach the bottom of the stack and run out of tasks, in any near future.

It’s not over - the work is changing, but not “done”.