A breed that stops existing

Most mornings I write a brief and hand it to an AI coding agent. The brief sets scope, constraints, and a definition of done. A few hours later the work comes back. Ten years ago that brief would have gone to a junior developer. Early in my career, that junior was me. The work I started with, and matured through, now goes to a machine. Which leaves a question I can’t put down: if the machine does the work that made me, where do the next people like me come from?

The usual way to say this is that AI takes junior jobs. True, and it misses the point. The output was the smallest part of junior work. You become a senior by making thousands of small decisions with low stakes, getting them wrong cheaply, and slowly building the judgment that later carries the big decisions. The boilerplate, the simple bugs, the grunt integration work: that was the training. It looked like busywork while it quietly built the judgment the title would later need.

So it feels like belonging to a breed that stops existing. But I’ve seen this before, from the other side. Starting out, I was the disruption. The seniors of the mainframe era watched their formation path disappear under exactly the layers I matured through, and they were certain real judgment couldn’t form without the discipline their scarce machines had forced on them. Every generation of seniors believes the path they climbed was the only one that could have produced them. The mainframe generation said it about mine.

The historical record is consistent: technology waves kill ladders, and mastery outlives them. Hand-loom weaving died as a lineage. So did assembly-language virtuosity. In both cases mastery re-formed around the new layer, climbed by people whose formation the old masters wouldn’t have recognized as formation.

Still, there’s an honest difference this time, and I won’t argue it away. Previous waves automated the floor, and juniors trained on the layer above it. This wave automates the rep-building layer itself. The thing being removed isn’t beneath the training ground. It is the training ground.

The best evidence that this can go well comes from chess. Engines were supposed to end human mastery. Instead, the generation that grew up with them is the strongest in history, reaching the top younger than anyone before, because the engine turned out to be a sparring partner with perfect feedback. More reps, tighter loops. Whether AI plays that role for software depends on something small and behavioral: whether the junior interrogates what the machine produces or just ships it. Same tool, two different formations.

Put the pieces together and the likely outcome is a missing generation. The lineage survives; it just skips a cohort. The market destroys the old ladder years before anyone designs the new one. Companies have already stopped hiring juniors. Nobody has yet built the path by which a junior becomes a senior without junior work. That gap produces a cohort hole, the way nuclear engineering hollowed out after the 1980s. The seniors of 2036 will exist. They’ll be formed differently, with judgment I might not immediately recognize as judgment: a feel for when the model is lying, orchestration instinct, system intuition built without ever writing the low-level code.

A cohort hole changes what my position is worth. The last people on a dying ladder, with a gap behind them, become scarce. The last COBOL programmers were paid better near retirement than at any point in their careers.

But scarcity is the passive reading. The people who climbed the old ladder are the only ones who know what the ladder was for: which judgments it actually built, which reps mattered, what the cheap failures taught. That knowledge is exactly what’s needed to design the new formation path, the sparring partner that does for software what engines did for chess. So the last climbers are the only ones who can draw the map for the next path, and the time to draw it is now, before the memory of what the old rungs taught goes cold.