As I reflected on our new self-image, I once wondered if I had ever written anything as simple as a decision tree entirely in pure Python. The answer was no. However, this is not at all alien to our role: data scientists have always imported the code that others have written in order to use it in their own code.
The agentic paralysis
The new path taken by our team had only taken the idea behind it to the next level. Almost a year after switching to “fully agentic”, my role as a data scientist feels more and more like what I originally imagined. What’s really at the core of a data scientist’s job is iteration: we build models, check the results, weigh things up, think about what we can adjust (or whether we should try a different approach), and then start the next iteration. Until we are satisfied with the result. By delegating programming tasks, we can iterate faster and more. It also makes it easier for us to reject a probably hopeless approach and start from scratch.
Given such a change in the role of the data scientist, a revised job description is not enough. Because this change is also accompanied by a change in working methods. At the core of this new way of working was for us to view Claude Code as an infrastructure – or even an employee – rather than a program that you open in the morning and close again in the evening.
AI agents need context – the tacit knowledge we have accumulated over years – and coding guidelines. They need smaller tasks that they can carry out perfectly and, overall, form the bigger picture. A lot of how we work now takes place in the concept phase – and when it comes to architecture. In other words, before any code is written. You have to understand that it’s not just about writing better prompts. It is essential to build an environment in which the agents function as an integral part.
However, after embarking on our new path, we quickly noticed something that felt counterintuitive: We were slowing down – the productivity gains were gone. From today’s perspective, this apparent paradox can be easily explained: we had to train the agents. They lacked context, so the generated code was incorrect – and you had to look very closely everywhere.
But with time and better context, the agents got better and better. After about three months, we had returned to our pre-change productivity levels and were becoming more and more confident in our new role. Last but not least, we were lucky: our boss gave us the time we needed. The change was not expected to immediately make us more productive. And that was a good thing.
