The Orchestrator

There is a widely held concern that generative AI is forcing developers to give up their understanding of how their systems work, and that we are heading toward a profession that ships code it does not comprehend.

Here is the thing: most developers already do not comprehend it.

We have always worked on top of systems we could not fully explain. The gap is not new; what is new is how much of the work now sits behind it. In a 2013 survey, 414 Java developers answered barely half the questions about what happens when their programs run against an upgraded library — and the self-described experts did only slightly better.¹⁰ Those developers shipped software that worked. The knowledge boundary has always been crossed by something other than understanding: tests, rollback plans, and the willingness to find out in staging instead of in production. Delegated intelligence does not remove that boundary. It widens it, and it removes the excuse — until now, you could say the system was too complex for one person to understand. You cannot say that about a system you personally delegated authority to.

That is not a failure. It is how engineering works: you understand the layer you are working in, you trust the layers below it to honor their interfaces, and when something breaks at a boundary you do not understand, you call a specialist. All of which is already delegation, and none of it is permission to be careless. It is the practical condition of building anything complicated, where nobody owns every layer and the skill is knowing which layer you must understand, which boundary you can trust, and who to call when that boundary fails.

So what generative AI changes is not whether developers understand their systems. It changes which layer they understand. A developer used to understand the application because they wrote it, and now the application may be substantially written by a model — which leaves the developer still needing to understand what it does, how it behaves, and where it fails, while no longer having read every line. That is a shift in where the understanding sits, not the loss of it. The job was never to understand everything. It was to understand enough to be responsible for what ships.

HQ 8 — Authored. I wrote the argument, structure, examples, and prose. AI tools were used for research assistance and light editorial polish; the voice, ideas, and substance are mine.