Fair. “It understands” is probably the emotional description, not the technical one.
The practical problem is that it can imitate understanding well enough to get a large project moving, then break down right where durable system memory and architectural consistency matter most.
So yes, it’s a tool. The problem is that it’s useful enough that “just don’t use it” is not a real answer, and broken enough that you eventually have to build around the gap.
Okay.But if I need three different framings just to converge on one coherent implementation, I’m still doing architecture recovery by hand. That’s a big part of why I started building something local around the codebase itself.
Yes, this is very close to what I meant.
They keep reconstructing the codebase from fragments instead of actually carrying the system forward, so you get re-implementation, duplicates, missing edge cases, and slow architectural drift.
Yes, exactly.
That’s why it feels so strange in practice. It can mimic understanding well enough to get you moving, but when the project gets deep enough, you find out it was generating plausibility, not actually holding the system in its context.
Mine goes a bit further than that. It structures the codebase first, feeds that into local agents, and then uses whichever model is on top with actual system context instead of making it guess from prompts.
I've laid the groundwork to do this locally, but the current crop of LLM agents just can't close the loop. Trying to get them to finish the job inevitably sends me right back into the spiral. Thinking i should do it manually.