AI Is Not Your Intern. And That’s the Problem.

When I first started using coding agents seriously, I made a familiar mistake. I treated the model as the main thing. I wrote better prompts. Then longer ones. Then, more carefully structured ones. Sometimes it worked beautifully. Sometimes it drifted. Same repo. Same goal. Different result. That inconsistency bothered me, because it revealed something deeper: The bottleneck was not raw intelligence. It was context. In AI Is Not Replacing You. It’s Reshaping How You Think , I wrote about how AI shifts the bottleneck upward. When generation becomes cheap, the differentiator becomes judgment, constraints, and systems thinking. That same pattern shows up in agentic engineering. The real leverage does not come from asking a model to do more. It comes from designing a better system around the model. The Mistake I Kept Making Early on, I was treating the agent like a very smart freelancer with amnesia. I would drop it into a task, explain the problem, and expect one well-written instruction