Coding agents are smart at implementation, but their architecture decisions tend to be conservative. Even when fed primary sources from cloud vendors, they'll suggest outdated patterns. Human intuition still wins on "we should use X instead" calls—especially when picking between modern vs legacy approaches. The agents optimize for safety over cutting-edge, which makes sense given their training data lag, but it means you can't blindly trust their system design choices. They're great executors, mediocre architects.