I originally developed this system when I was first building applications on Replit, early in my Assisted Software Development days. Anyone familiar with platforms like Replit, Lovable, or other modern coding agents knows that while they are constantly improving, they have a bad habit of getting stuck. If you don’t feed them a flawless, highly detailed implementation plan from the start, they’ll do bizarre things.
For example:
They will entirely forget to build a database into a system that strictly requires one.
Or,
They will build two separate databases, fill one with data, but configure the system to read only from the empty one.
As you may have guessed, both of these things happened on early builds of mine. And when things go wrong, you end up trapped in never-ending debugging loops where the agent simply cannot figure out how to fix its own mistakes.
I’ve seen this happen repeatedly, even with custom agents. I actually built a debugging function into My Digital Employee, though I haven’t developed it extensively because I don’t want my primary agent acting strictly as a programmer. I just wanted it to have the baseline ability to problem-solve its way out of a corner if it hit a wall. Nevertheless, whenever I allow it to diagnose and “fix” itself, it gets trapped.1
Similarly, when a development platform like Replit gets stuck in a debugging loop, it’s surprisingly difficult to break out if you don’t possess the technical knowledge yourself. If you can’t read the code to see exactly what is breaking and point the agent in the right direction, it will just loop indefinitely, leaving you stuck and quickly burning through the set monthly token usage.
During one of my very first projects with Replit, this exact scenario happened, and I was forced to abandon the entire project.
While I have since successfully built something in Replit, it will be the last time I use it2. In my current workflow, I prefer to
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.