Engineering discipline

Has anyone else felt like they’ve designed what seems like a solid architecture using AI tools, and then harnessed it through coding agents like Claude Code/Codex — only to realize the project is moving so fast that you’re starting to lose comprehension of what’s actually being built? I’ll be honest: I don’t really care about every line of code being written. I care about the architecture, the engineering decisions, and whether the system actually works. But that’s where I’m struggling. How do you maintain engineering discipline when AI can generate and modify code much faster than you can realistically review and understand every change? How do you make sure you’re not just building AI slop on top of what initially looked like a great architecture? I’ve been thinking about loop engineering as a solution, but I’m starting to feel like it isn’t enough. We build → observe bottlenecks → tweak the architecture → build again → discover new bottlenecks → repeat. At some point, the architecture itself keeps evolving faster than your mental model of the system. So I’m curious about people actually building serious systems with coding agents: How do you maintain engineering discipline and architectural integrity when the code generation is moving faster than your ability to comprehend the entire codebase? And am I misunderstanding loop engineering here? Is continuously iterating on the system actually the right answer, or is there another discipline/practice that keeps agent-assisted development from turning into AI slop? Would genuinely love to hear from people who are dealing with this in production, not just building demos. I do know the obvious answer is to slow down, read the code, and build incrementally. But when the whole point of these tools is to massively accelerate the feedback loop, is there a better engineering practice that lets us keep that velocity without sacrificing understanding and discipline?