Essays/linkedin/ scientific python programming with ai agents

<p>The <strong>sweet spot</strong> for AI-developed software lies in single-purpose programs, not the next "unicorn" company. But how can we incentivize and teach its use?</p> <p>A few weeks ago, I was dusting off materials for the Python workshops I used to organize. It occurred to me to test what would happen if I used an AI agent to develop a program similar to what students create throughout the course.</p> <p><strong>And it just worked.</strong></p> <p>The caveat is that I knew exactly what I needed and how I wanted it. I knew which libraries to use, how the measurements should happen, and where the potential bottlenecks were. I spent a couple of hours on something that would have easily taken a full day if I had written all the code from scratch.</p> <p>People overstress the idea of "production code." The reality is that the software I wrote would be used only by me—or, in a larger context like a lab or company, perhaps by a handful of people. It isn't exposed to millions of users on a website, nor does it handle sensitive data like payments or passwords.</p> <p>I believe that is the true sweet spot for the current state of AI tools: <strong>single-purpose, extremely customized programs.</strong> Why settle for a generic piece of software when you can have one that does exactly what you need, exactly how you need it?</p> <p>In my case, AI is enhancing my workflow; it is doing the same work I would have done, just faster. The next step is figuring out how to teach this new way of working. How much coding training is necessary for someone to become fluent in using AI tools?</p> <p>Last week, I wanted to visualize some high-dimensional flow cytometry data. I had done it in the past, but I wanted to add interactivity to the plots—selecting ranges, performing cross-correlation analysis, and so on. In my mind, the goal was clear; what blocked me was knowing how painful it is to program interactivity: handling clicks, checking ranges, extracting subsets, updating plots, and managing exceptions.</p> <p>My feeling was that if I could explain it clearly, I would get what I wanted.</p> <p>Yet, my impression remains that someone who has never developed a desktop application would likely not benefit from this approach. There is still a "knowledge jump" to be made.</p> <p>I’m curious: how many single-purpose programs are being created with AI tools? What are your challenges? Have you ever had to explain this process to a colleague?</p>