That is the question of our age.
It started with litter.
Back when I first created the mathematical model behind Litter Logger—an app that logs litter and, more importantly, shows where it’s being picked up by everyday community heroes—I stumbled on an idea. That address calculation model had a hidden power: it could aggregate enormous amounts of data points incredibly fast.
That’s when the thought hit me: if it worked so well in one app, what could it do if I built an entirely new one around it?
Fast forward to the present, where AI and “agentic” tools are quickly carving out space in the developer’s toolkit. Naturally, I couldn’t resist the question: just how good are these systems? Could they actually build an entire app—not just pieces, but the whole thing?
So, I put the Location Address App to the test.
Building with Agentic AI
Most of this app was developed through GitHub’s agentic services. And let me tell you—it’s been eye‑opening. At times, it’s brilliant. It implemented parts of my design in ways that genuinely improved on my original approach. Even better, it pulled off features I haven’t yet learned myself, like animations.
But the process also exposed real shortcomings. For AI agents to work smoothly, you absolutely need a Git repository and frequent commits. Things will go wrong. Sometimes it stumbles over the simplest requests, or just gets confused. Those moments taught me the value of giving the system a proper structure and safety net to work with.
And then there’s the code it produces. While functional, it tends to be far more inflated than a developer would normally write, and it’s not great at re‑using components. It’s almost like it writes in bursts of inspiration, but doesn’t always tidy up afterward.
Lessons Learned
In the end, what I got was better than I expected. My design was improved in subtle, smart ways, even if it came with trade‑offs. The inflated code and lack of reuse might frustrate seasoned engineers, but the fact remains: it worked.
It also gave me a glimpse of what’s coming. These limitations don’t feel permanent—they feel like growing pains. Right now, models can only partially digest a codebase. Sometimes they see the whole picture, other times just fragments. As they get smarter, sharper, and more efficient, so will we.
Final Reflection
To AI or not to AI—that was my question going in. After this journey, my answer is clear: today’s tools aren’t perfect, but they’re already powerful enough to open new doors for builders. And tomorrow? They’re only going to get better.
The real question now is: what will you build when they do?
Leave a Reply