In December 2025 the LLMs from Anthropic and OpenAI passed a threshold. They were no longer an interesting curiosity for professional software engineers. They were the tools that completely changed the work as a software engineer.

The change happened fast and rushed through my software engineering org like a shock wave.

I was able to witness first-hand how this change unfolded. In January, I read the first blogs on Hacker News, where AI sceptics had changed their mind about the usefulness of AI. Suddenly they were overly optimistic.

Then I had my own first-hand experience with the frontier models and was very impressed. My first session with a capable AI Agent in my IDE was eye-opening. As a manager I don’t work on production code anymore. But I still had my own side projects. I spent a few hours each week working on small coding tasks. With the AI in the IDE I was so powerful, that I was able to complete several projects in a few days.

From that moment on, I started actively pushing for AI adoption in our engineering org. I scheduled sessions with each team and together we shared our experiences and showed each other what was possible.

I was surprised to see that even after a few months after the release of the new frontier models, people still hadn’t used them. It sometimes took encouragement from me and my colleagues to get people to use them. But sooner rather than later everyone had their heureka moment and now we live in the new reality of software development.

Looking up problems no longer takes a Google search ending up on Stack Overflow. Asking the agent in your IDE directly gives you the specific answer for your codebase.

  • Typing code is not a must anymore. Of course, sometimes it is necessary and faster than prompting the Agents, but if you have to implement a bigger chunk of code, it’s much more convenient to simply let the agent implement the change.

  • Being blocked by not knowing where to start is also not a problem anymore. You simply give the agent a very rough idea of what you want to achieve and it takes the lead and starts pushing to finish the task.

  • It feels different to be a software engineer. In the past, it was a craft where it mattered to structure code well and find elegant abstractions for the implementation you were working on. It was important to memorize concepts and syntax. You needed a lot of experience to dive into a large new codebase and grasp the key concepts of the implementation. Now you delegate a lot of the details to your AI Agent.

  • AI Agents are not only reading code. They are also building and executing the project. Now automation is even more important than before. If the Agent cannot build your project and run it, it is far less powerful. I experienced the value of an agent beeing able to built the project, run it and also operate it via an API. In this way the agent was able to really test and debug on its own. In this way the agent can close the loop and go all the way from requirements engineering to testing the implemented feature. Sometimes this feels like cheating.

  • There is also the other side. It occurs when the AI is stuck. When it is running in circles and not understanding what you really want to achieve. It burns a lot of tokens, makes huge changes to the codebase but does not achieve your goal at all. These cases really suck.

So what does this mean for the software engineers out there? I honestly don’t know. But I still see a lot of demand for skilled people who can solve complex problems with the help of AI.

And the real bottleneck never was producing code. It always was finding and solving the problem of the customer with the right code. That problem is still hard and needs human intelligence to tackle.