Last month I sat through a hiring loop for what was, on paper, a mid-level role. Five candidates, all with three to five years of experience, all employed, all coming from real companies you would recognize. The screen was a small refactor — a hundred-line function that needed splitting, with a quiet bug tucked into the conditional logic.
Four of the five candidates produced clean refactors. Their code was better organized than the original. The bug was still there in all four refactors. None of them noticed.
The fifth candidate noticed in eleven minutes.
I have been thinking about that ratio — 1 in 5 — for six weeks. And I have come to a conclusion that I do not love: the senior engineer is becoming an endangered species, and almost nobody is talking about why.
How seniors are actually made
A senior engineer is not a job title. It is a person who has, over years, been on the wrong end of enough decisions to develop a working model of how things break. Not in theory — in their own hands, on their own watch, with users yelling. That kind of knowledge cannot be downloaded. It is grown, slowly, the way coral grows.
The way it has historically been grown is: a junior is given small, real work. They do most of it wrong. They are corrected, gently or not, by someone more senior. They do the next thing slightly less wrong. After about two years of this, they are a mid-level engineer — meaning, roughly, they can be left alone with a medium-sized problem without immediate disaster. After another three or four years, they are a senior — meaning they can be left alone with a large problem and will probably notice the disaster before it happens.
The total pipeline is about five to seven years. Seniors are a depleting stock unless you are continuously refilling the input.
In every industry that has ever had a senior class — surgery, law, aviation, plumbing — this pipeline has been the load-bearing assumption. It is the assumption you cannot see because it is always quietly true.
What's actually changed
The mistake most people make when thinking about AI and engineering is to think it's reducing the demand for seniors. It isn't. It's doing the opposite. AI generates more code than ever, which means more code needs to be reviewed than ever, which means seniors — who can tell good code from bad — are more valuable than they have been in a decade. Compensation for genuinely senior engineers is, by every signal I have access to, going up.
The thing AI is reducing is the supply of juniors who can become seniors. Not the supply of juniors who can be hired — there are still plenty of those — but the supply of juniors who, given five years, would have grown into the people we now badly need.
The mechanism is simple and almost unnoticed. The work that was the way in — the small bugs, the simple refactors, the unglamorous validation logic, the boring CRUD endpoint, the test scaffolding — has been swallowed by tooling. A junior in 2020 spent two years writing that work and developing taste in the process. A junior in 2026 watches the model produce the same work in fifteen seconds and is then asked to review it.
Reviewing is a senior skill. We are asking juniors to perform it on day one.
The factory floor that used to make seniors is still running. It just stopped accepting raw material three years ago. Right now we are still shipping seniors out the back. We have not yet noticed there is nothing coming in the front.
The latency that hides the problem
The thing that makes this hard to see is the five-to-seven-year latency. The seniors of 2026 are people who came up before this shift. They are still being produced from the cohort that started juniors in 2019, 2020, 2021. That well is not yet dry. It is, in fact, still flowing well enough that hiring managers can mostly find what they need, if they look hard.
Sometime in 2028 to 2030, that well will run out. The cohort that started juniors in 2024 and 2025 will, on average, be less prepared to step into senior roles than any cohort in the last twenty years — not because they're worse people, but because the apprenticeship that used to come with the job has been taken away from them and replaced with a tool that does it for them.
This is the part where the column-writers usually pivot to and that's why we need to invest in education or some other comforting abstraction. I don't think education fixes this. The thing being lost is not knowledge — it is the failure-on-real-systems that produces taste. You cannot teach that in a bootcamp because you cannot fake the production incident at 4am.
The companies that will survive this
The interesting question is what individual companies will do about it. Most won't do anything. They will keep hiring AI-fluent juniors, getting two or three years of accelerated output from them, watching them plateau, and then complaining at conferences that "you can't find a real senior anymore."
A few companies will do the contrarian thing and deliberately absorb the cost of training juniors slowly, on real work, even when AI could do that work cheaper. They will pay a junior to refactor the hundred-line function by hand. They will let a mid-level engineer review the PR carefully. They will swallow the productivity gap on purpose, year after year, because they understand that the cheapest way to acquire a senior in 2032 is to be paying for one to grow in 2026.
These companies will look, from the outside, like they are run by people who don't understand how to use modern tooling. They will be the ones who still have engineers in 2032 who can spot the bug in eleven minutes.
There is a precedent for this. Most of the surgical residencies in the US do procedures by hand that could be done faster and more cheaply by robotic assistance — and they do them by hand on purpose, because they know that a surgeon who never learned the manual skill is a surgeon who cannot recover when the robot is wrong. Every senior engineer I know has, somewhere in their career, been the human who recovered when the robot was wrong.
A small admission
I don't know what to do about this at the industry scale. I am one person who hires occasionally and who writes a journal that twelve people read. The problem is bigger than any one team's policy.
But at the personal level, I have started doing one thing: when I onboard a junior now, I deliberately give them the boring tasks. The validation logic. The test scaffolding. The thing the AI could do in a minute. I sit with them for an extra forty-five minutes while they do it the slow way, and I correct them, and I watch them get faster. It is, in the short term, an enormous waste of my time and the company's money.
I think it is the only thing I know how to do that has any chance of helping.
The fifth candidate, the one who noticed the bug in eleven minutes, came up before 2021. I asked her how she had developed the instinct. She thought about it for a long time and said: I spent two years fixing other people's bad code, in a job nobody would give a junior today.
She did not say it like an achievement. She said it like an inheritance she had not earned, and could not pass on.