The shift from "writing code" to "orchestrating agents that write code" is real, but I think the underappreciated skill is knowing when NOT to delegate to the agent. The best agentic engineers I've seen treat AI like a junior dev — great for boilerplate and well-defined tasks, but you still need a senior engineer's judgment for architecture decisions, error handling strategies, and understanding the business context behind technical choices. The "death of coding" framing misses this: what's dying is typing speed as a bottleneck. What's rising is the ability to decompose problems, evaluate AI output critically, and design robust error recovery flows.