Nobody Is Minting Mid-Level Engineers Anymore

A blog post by Florian Herrengt hit the Hacker News front page this week with a blunt thesis: AI is removing the middle class of software engineering. Strong engineers get more valuable because AI multiplies them; weak engineers get harder to justify because AI-amplified mediocrity now produces damage faster than anyone can review it. The middle of the pay curve — the reliable mid-level engineer who turns tickets into pull requests — gets squeezed out.

The diagnosis resonates because everyone has now lived some version of his anecdotes: the 24,000-line PR nobody can meaningfully review, the junior who answers "where does this data come from?" with "let me ask Claude." But the post gets one important thing backwards, and the correction matters for anyone planning a career or a hiring budget. The middle class of engineering isn't being fired. It's being un-minted — and that's a slower, quieter, and ultimately worse problem.

What the employment data actually shows

The best measurement we have of AI's effect on the software labor market is the Stanford Digital Economy Lab's "Canaries in the Coal Mine" study, which analyzed payroll records from ADP covering millions of workers. Its findings are precise in a way the discourse usually isn't: employment for 22-to-25-year-olds in the most AI-exposed occupations — software development prominent among them — fell roughly 13% relative to less-exposed peers since late 2022. Meanwhile, employment for workers in their mid-to-late thirties in the same occupations kept growing.

Read that again with Herrengt's thesis in mind. The measured damage so far isn't in the middle of the ladder. It's at the bottom. Mid-career engineers are, for now, the beneficiaries: they're the ones companies trust to wield the code-generating firehose. The Stanford team's second finding is the one that should worry you more, though — the decline isn't driven by layoffs. It's driven by a collapse in hiring. Nobody is getting marched out; the door just stopped opening.

That's the mechanism by which the middle class actually disappears. Mid-level engineers aren't hired off the street — they're grown. They're the juniors of three years ago who shipped, broke things, got reviewed, and absorbed a system's tribal knowledge. If the industry stops hiring 23-year-olds in 2025 and 2026, it doesn't feel the shortage today. It feels it in 2029, when the seniors it's leaning on start leaving and there's no cohort behind them. The middle isn't being removed by AI. It's being cancelled in advance, one unposted junior req at a time.

The productivity story is shakier than the hiring story

Here's the uncomfortable part: the belief driving those hiring decisions rests on productivity claims we can barely measure. METR ran the most rigorous experiment to date — a randomized controlled trial with experienced open-source developers working on repositories they knew well — and found that early-2025 AI tools made them 19% slower, even as the same developers estimated they'd been sped up by 20%. When METR tried to rerun the study in 2026, it couldn't get a clean read: developers refused to participate in the no-AI condition, and 30–50% quietly withheld tasks they didn't want to do by hand. The tooling has almost certainly improved since the original study, and METR says as much — but the honest summary is that we're restructuring an entire profession's career ladder based on perceived productivity gains that resist measurement.

That doesn't make Herrengt wrong about the day-to-day dynamics. His sharpest observation is about review economics, and it holds regardless of what the true speedup number is: AI removed the natural speed limit on code production but did nothing for the speed limit on human verification. When one engineer can generate a weekend's worth of plausible-looking changes that would previously have taken a month, the scarce resource in the organization stops being "people who can write code" and becomes "people who can be accountable for code." That's a real repricing. It just reprices judgment, not seniority per se — and plenty of mid-level engineers have judgment.

What to actually do about it

If you're a mid-level engineer, the move isn't to out-prompt the tools — Herrengt himself argued elsewhere that vibe-coding skill takes weeks to acquire and confers no moat. The durable position is accountability for a system: be the person who owns the payment pipeline or the auth layer end-to-end, who can say "no" to a 24,000-line PR and articulate why, and whose review is the one that unblocks merges. Verification skill — reading diffs fast, writing the test harness that catches the agent's confident nonsense, running the incident when it slips through — is the thing the firehose made scarce. Concretely: cap PR size in CI, insist on design docs before generation sprees, and get good at the archaeology of systems you didn't write, because AI-authored codebases are all systems nobody wrote.

If you run a team, the Stanford data is a warning about your own five-year pipeline, not just the industry's. Junior engineers plus AI tools are cheaper to grow into your 2030 senior bench than 2030 senior salaries will be. The companies quietly killing entry-level reqs are making a locally rational, globally ruinous trade — and creating an arbitrage for anyone willing to keep training people.

Herrengt's post deserves its front-page run because the bifurcation he describes is real. But "AI removes the middle class" implies a purge that the data doesn't show. What the data shows is scarier: an industry eating its seed corn while measuring the harvest with a broken scale. The middle class of software engineering won't be removed. It just won't show up — and we'll spend the early 2030s paying whatever the surviving seniors ask.