I think it's more of a "how you use it" thing, but you're definitely right that AI agents can't design systems properly.

Some people I know have produced way more code, removed tech debt, and all without introducing any bugs since they started using AI. That's because they're not using it to do anything beyond their skillset, understand everything it's doing, and are using it to catch mistakes they otherwise would have made. Other people are using it without reviewing the output, or are using it to try and do things beyond their skillset, and that's how you end up with infinite tech debt and a whole host of bugs.

Personally I've recently started heavily using the AI code review bot we have at our company, both for my own code and other people's. While 50% of what it says is hallucination or wrong, that's not an issue because I know it's wrong or a hallucination so can just tell it no and to focus on other things, like catching bugs or issues that most reviewers would just glance past, and also gives you a rubber duck that talks back.