The pyramid is narrowing at the base
Brynjolfsson and his Stanford Digital Economy Lab co-authors (Chandar and Chen) put a number on something a lot of us had been feeling. In their 2025 paper, developers aged 22–25 have lost about 20% of employment since late 2022. Seniors in the same occupation stayed flat. The early-career hit shows up across AI-exposed occupations at about -13% versus trend; software is one of the sharper cases.
The number is news. What it means operationally is more uncomfortable, and that's the part I want to sit with.
The data and the short read
The short read writes itself: "juniors are going to disappear". And in a narrow sense, the data says exactly that. The 22–25 cohort isn't getting hired into software the way it was three years ago. The agents absorbed the work, the work absorbed the headcount, the headcount didn't come back.
The trouble with the short read is that it stops at the symptom. It treats "juniors disappearing" as a hiring-budget problem you solve by either (a) accepting it and saving the salary line, or (b) holding the line and hiring them anyway out of some combination of optimism and policy.
Neither response addresses the structural thing underneath. Which is that the pyramid isn't just narrowing at the base. The pyramid had a specific function at the base that nothing else currently performs. And the function is the part I keep getting stuck on.
What the pyramid did when it worked
The bounded-scope junior role — writing CRUD endpoints, fixing tagged bugs, completing small features, doing the ticket queue — was never just labor at the bottom of an org chart. It was the practical classroom where seniors got made.
Ten years of bounded-scope work, accumulated, did three things you can't easily get any other way. You internalized patterns — what production systems actually look like, where they fail, what kinds of mistakes you can recover from and what kinds you can't. You built intuition for what breaks — the muscle memory of "this is going to bite us in three months" that nobody can articulate but every senior has. And you developed mature technical vocabulary — the ability to talk about systems with people more senior than you, in their language, because you'd spent enough hours in the substrate to earn the words.
That ten-year arc was the pipeline. Not a metaphor — a literal sequence of bounded-scope problems that produced the cognitive equipment of a senior engineer.
And here's the part I keep returning to: those bounded-scope roles are exactly the ones the agent absorbs best. The CRUD endpoint. The tagged bug. The small feature. The work that was "good for juniors" because it was contained, well-specified, and rewarded persistence — that's the agent's natural territory now. The practical classroom and the agent's optimal workload are the same set of tasks.
The pipeline gap
Which means today's senior cohort — the engineers in their late 30s and 40s who came up through that pipeline completely — is the last cohort the old system trained end to end. They had the full ten years of bounded-scope work before the agent showed up. The juniors coming in now don't have that practical classroom available, because the practical classroom is being done by the agent.
And nobody has a replacement for the pipeline. Not academia — university CS programs were never the pipeline; they were the entry point to it. Not bootcamps — bootcamps were a faster entry point to the same pipeline. Not companies — most companies don't have a deliberate training architecture for what comes after "hired". They had the pipeline as exogenous infrastructure. They consumed it without paying for it.
The few signals I see of people trying to build a replacement are tentative. Russinovich and Hanselman, in their forthcoming 2026 CACM piece, propose modeling something like medical preceptorship — the explicit pairing of trainees with experienced practitioners on shared cases, with deliberate handoff of judgment over time. It's a real idea. It's not yet a solution. It's the kind of proposal you make when you've recognized the gap but the architecture is still being designed.
The open question
So here's the question I genuinely don't have a tidy answer to.
If in 7 to 10 years you need Mixer-fluent seniors — practitioners who can hold product/architecture/code/QA simultaneously and govern a fleet of agents from that integrated view — where will they come from, if you're not training juniors today?
The honest options as I see them are three.
One: redesign entry-level roles as "agent supervisors" — bringing juniors directly into Pillar 2 work without the bounded-scope production decade. The question nobody has answered is whether this actually produces the same cognitive equipment, or just produces people who are good at supervising agents without having the substrate-level intuition that made seniors valuable in the first place. I don't think we know yet. It's a hypothesis being tested in real time.
Two: accept a structural shortage. Let the pyramid narrow, let the senior cohort age out, and let the market price the shortage. This works for individual companies in the short term — buy the seniors that exist, pay what they cost. It's a tragedy of the commons in the medium term, because the seniors-that-exist pool is finite and not renewing.
Three: assume "senior" as a category shifts so much in the next decade that this projection is wrong — that the thing we need on the other side isn't really continuous with what we mean by senior today, and the pipeline question is therefore reframed rather than answered. This is the most epistemically modest option and also the most likely to be true in some form. It doesn't help anyone making 2026 hiring decisions.
I don't know which is the answer. I distrust anyone who tells you they do. The data is too fresh, the experiments too few, the timeline too compressed.
What I can name
What I can name is the asymmetry of the cost.
The CIO not thinking about a 5-year pipeline today is mortgaging their organization's technical capacity in a way that won't appear on any current quarter's report. The cost shows up later, when the seniors they're depending on retire, leave, or get poached, and the bench is thin in a way that wasn't visible while it was being created.
The CTO replacing juniors with agents without designing the alternative pipeline is consuming senior capital that doesn't renew. Every senior who reviews agent output, who carries the setpoint, who holds the integrated mental model, is drawing down a finite pool. The pool was built by years of bounded-scope work that isn't happening anymore. There's no inflow.
The defensive move — not the bold one, the defensive one — is explicit preceptorship. Pair juniors with seniors on shared projects, not separate tasks. Make the senior's reasoning visible. Have the junior do the parts the agent doesn't do well yet — context-building, stakeholder translation, the messy human-in-the-loop work — under the eye of someone who's been through the full arc. It's slower than just hiring fewer juniors. It produces fewer hours of code per quarter. It also produces engineers who exist on the other side of the transition.
I'm not selling it as the answer. I'm saying it's the move that doesn't burn the substrate, which feels like the minimum bar.
Is your org hiring juniors this year? In what role? Curious about the real experiments more than the opinions.
#Hiring #FutureOfWork #Engineering #Career #TechLeadership