Is anyone else noticing that people building AI agents are forgetting basic software engineering?

I've been reading a lot of posts lately from people struggling with agent memory, prompt injection and goal hijacking. These are real problems, and I'm not dismissing them. But what I keep seeing is that the basic software engineering approach has gone missing. We never trusted what a client put in a form. We designed the data model first, validated on both the client and the server, and handled injection at several layers. We isolated processes and data so that one compromised service didn't leak everything. We used role-based access with short-lived tokens. Most of those same basics would solve a lot of what people are hitting with agents. Take memory. If you don't know what you want to store and how it should influence the agent, you can't design a data structure for it. So everything gets stored as free text because "it's an LLM," and of course you end up with injections. I think part of the cause is vibe coding. Even if you vibe code a full-stack app in one shot, the AI will probably cover the end-to-end flows. It won't handle deep validation, database indexing and the rest. You add those before you go to production. Agent workflows are the same. A vibe-coded agent won't address the cross-cutting concerns for you. You still need to know software engineering, design principles and architecture to tell it what you actually want. What are your views?