Laurie Voss, who co-founded npm, recently published a post whose title doesn’t hedge: “AI has torched the market for junior programmers”. The numbers deserve the click. Developers aged 22 to 25 are down 19% from their late-2022 peak (ADP payroll data), entry-level postings are down 28%. And total developer employment is up 4.4%. The profession is fine. The entry point is on fire.
Every time a headline says AI is coming for programming jobs, I ask which programming job, specifically. A software engineer’s job has always been two jobs wearing one title: deciding what code should exist, and then writing it. AI got good at the second job first. Fast. The junior end of the market was built almost entirely on that second job, and Voss’s split, juniors down 19% while the total is up, is what it looks like when only that half goes.
- 01 The job that is disappearing is code written to a spec someone else already decided. Most junior roles were built on exactly that.
- 02 AI got good at execution, the easier half of the job. Judgment, deciding what should be built, still sits with the people who already had it.
- 03 Remove the execution work and you remove the accidental curriculum juniors learned judgment from. Hiring and mentoring have to move with it.
- 04 This is a warning for anyone, junior or senior, whose value was mostly throughput rather than judgment.
Two jobs, one title
Job one: work out what code should exist. Right fix, or symptom? Does this edge case matter, or is it noise? Should we build this at all? Job two: turn that decision into working code. The ticket, the spec, the acceptance criteria, executed correctly.
Most junior roles were quietly “do job two, under supervision, until you have absorbed enough of job one to be trusted with it”. Not a criticism of juniors. It was a reasonable way to grow judgment: years of execution while watching seniors make the calls around you. It just assumed the execution step would always need a human doing it slowly enough to learn from.
The reframe: execution and judgment were never the same skill
An AI agent can turn a well-written ticket into a finished, reviewable change in minutes. That automates the mechanical half of the job, the half the junior market was built on. The harder half, deciding what should be built, still sits almost entirely with people who already had the judgment before the tools arrived.
Before: Most of the queue is "turn this spec into code." The rare judgment call gets escalated to a senior.
Now: Most of the queue is "decide whether what the AI wrote is right, and whether the spec was even the right thing to build." The execution step barely touches a human at all.The queue inverted. The work that used to fill a junior’s week (and, not incidentally, used to teach a junior how a senior thinks) is the part an agent now does before you have finished reading the ticket.
What this actually changes for hiring and mentoring
Voss frames the old model as an apprenticeship pipeline that AI has now broken. Juniors absorbed judgment by writing code under senior review, for years; the writing is the part that’s gone. I would add the view from the hiring side. Most engineering orgs have no deliberate mechanism for teaching judgment on its own; it was always a byproduct of years of execution work. Remove the execution work and you have removed the accidental curriculum, not just the busywork.
On my own team the shift has already happened, and the honest version is less dramatic than the headlines. Our hiring pipeline used to span mid-level to senior engineers; the guideline now is senior only. We also screen for engineers who are AI-forward rather than AI-sceptical, written into the hiring guidelines rather than left to interviewer taste. And AI gets less credit than you might think: we were already drifting that way before the tools got good. AI didn’t start the shift. It made it permanent. I cannot see us going back for the foreseeable future.
What I would hire and mentor for now is the question the spec never answers: can you tell me whether this is worth building, and why. That is a harder thing to teach and a harder thing to interview for, and nobody, including me, has fully redesigned the on-ramp for it yet.
None of this makes junior engineers obsolete. The cheap, plentiful training task they used to learn on is gone, though, and the only thing left to hire and mentor for is judgment — which is exactly the thing most engineering orgs never had a deliberate way to teach.