The middle is falling out of software engineering

What happened

A post by Florian Herrengt titled *AI is removing the middle class of software engineering* hit 870 points on Hacker News this week, and the comment section did what the comment section always does when a piece touches a nerve: it kept going for 1,200+ replies without cooling off. The argument is simple. AI coding tools haven't replaced *engineers* — they've replaced the *kind of work* that mid-level engineers were doing. CRUD endpoints, glue code, a React form wired to an API, a migration script, the thousandth internal admin page. That work still exists. It just doesn't need a human with three to seven years of experience anymore.

Herrengt's frame is economic, not technical. Senior engineers get a productivity multiplier from AI; juniors get a cheap on-ramp; the middle tier gets compressed because the tasks they were priced for have collapsed in value. The people who used to charge $180k to convert a Figma into a working feature are competing with a Claude Code session that a staff engineer runs while eating lunch. The people who charge $400k because they can hold a distributed system in their head, decide what *not* to build, and unblock ten other engineers — they're fine. For now.

The HN thread is unusually candid. Multiple commenters describe getting laid off from mid-level IC roles in 2025 and finding that the market for "5 years, full-stack, ships features" has quietly dried up, while staff+ roles and true junior hires (cheap, moldable, no bad habits) are still open. One commenter, a hiring manager, put it bluntly: "I don't need three mids anymore. I need one senior with Cursor and one junior I can shape."

Why it matters

This is the first labor-market pattern from AI coding tools that looks structural rather than cyclical, and it inverts the story the industry has been telling itself for two years. The reassuring narrative was: *AI replaces the boring parts, humans move up the stack, everyone wins.* That works if the ladder is intact. It isn't. The middle rungs of the ladder were the boring parts.

Juniors historically became mids by doing exactly the work that AI now does in thirty seconds — writing the tenth CRUD endpoint, debugging a flaky test, refactoring a component. That was the apprenticeship. Remove it and you get a generation of engineers who can prompt but can't debug the output when it's wrong at scale, because they never built the intuition. This is the same critique senior devs have been making about Copilot-only workflows since 2023, but it's now colliding with an economic incentive: companies don't have to pay for the apprenticeship if the output shows up anyway. So they aren't.

Compare this to the earlier automation waves that didn't hollow out the middle. When compilers replaced hand-written assembly, the assembly experts moved up into systems roles and a new mid-tier (application programmers) formed under them. When AWS replaced sysadmins, the sysadmins became SREs and a new mid-tier (platform engineers) formed. Each wave created a new floor. This wave is different because AI operates at the level of the abstraction itself — it writes the code, reviews the code, and increasingly runs the code. There's no new mid-tier forming underneath because the abstraction ate the mid-tier's job description.

The counter-argument, made forcefully in the HN thread, is that this is temporary. AI still hallucinates. It still writes code that compiles and passes tests but does the wrong thing in production. Someone has to catch that, and "someone" is the mid-level engineer with production scars. The problem with this counter is that catching AI mistakes is a *senior* skill, not a mid one. You need to know what "wrong" looks like at the architecture level, not the line level. A mid-level engineer whose entire career has been shipping features to spec doesn't have that muscle — they've been *given* the spec, not written it.

There's also a compensation math problem hiding here. The mid-tier at $150-250k was priced on a specific bundle: enough experience to work independently, not so much that they cost staff+ money. AI eats the "work independently on well-scoped tasks" part of that bundle. What's left — judgment, taste, architectural instinct — is what seniors are already paid for. So the market re-prices: either you're worth staff money because you have those skills, or you're worth junior money because you're being trained, and there's no coherent middle price point anymore.

What this means for your stack

If you're a mid-level engineer reading this: the honest move isn't "learn AI harder." Every mid is learning AI. Prompt fluency is table stakes now, not a moat. The move is to deliberately take on the work that doesn't fit in a context window: cross-team coordination, incident command, migration planning, deprecation, mentoring the juniors who now report to you, owning a system end-to-end for two years. That work compounds into seniority. Feature-shipping doesn't anymore.

If you're hiring: the barbell strategy commenters describe (one senior + one junior instead of three mids) is real, and it's probably rational in the short term, but it has a five-year problem. Your juniors are going to hit the mid-level cliff too — they'll need to become seniors without ever being mids, and no one knows how to run that apprenticeship yet. Companies that figure out a deliberate senior-development pipeline in the AI era will have a decade-long hiring advantage. Everyone else will be bidding against Meta for the same 500 staff engineers in 2030.

If you're a manager: audit your team's actual work, not their titles. If your mids are doing what AI now does, that's a leading indicator, not a stable equilibrium. Reassign scope now — real ownership, real ambiguity, real customer contact — or the layoff conversation is already scheduled, you just don't know the date.

Looking ahead

The optimistic read is that this is 1995 all over again: the web wiped out a swath of desktop-software roles, panic ensued, and five years later there were more software jobs than ever, just distributed differently. That could happen again. But the pessimistic read has more evidence right now: hiring data, layoff patterns, and the plain fact that GPT-5-class and Claude-Sonnet-class models can already do a competent junior-to-mid feature ticket end-to-end. If you're betting your career on "the middle will come back," you're betting against the observable trend. Better to bet on climbing out of it — while there's still a ladder to climb.