Every time you need to make an update, you need to bring up the old context, or otherwise get the AI up to speed, which especially if you're using one of the frontier models could be a significant financial drain long term.
You don't get the same dopamine hit too, because you're just making boring updates to something which you threw together in 5 minutes with zero effort. The time and financial cost of building all this stuff may have been better spent on one, good, properly architected project.
Maintaining the project manually also assumes you can quickly understand the codebase which has been produced, otherwise you're completely dependent on Anthropic and them maintaining prices which you can afford. Bearing in mind that as you add new features, the cost of getting the LLM to understand the project increases, right? I might have a naive perspective here.
All that being said - sometimes there really are one-off niche things that are just for personal use that you do continue to use long term. Usually the simpler stuff where you can easily grasp the codebase at a quick glance. It's also great for debugging back and forths.
Personally I just run my local setup with a bunch of MCP stuff and the primary way it helps me is to keep me functional and on task. In some ways it's good if the AI can supervise you as opposed to you supervising it - at least from an ADHD perspective.
It's an interesting idea for sure, I like this article and agree with it.