Opinion: AI Is Removing the Middle Class of Software Engineering

In an essay published on his blog, engineer Florian Herrengt argues that AI has removed the “speed limit” that once kept bad engineering decisions from compounding quickly. His scenario: it is an ordinary Monday, you are the senior engineer responsible for code quality, and you open seven pull requests to review — the first spanning tens of thousands of added lines, delivered with an AI-written description.

Fast to write, slow to understand

The essay's central point is that, to the untrained eye, this way of working appears to function: you pull the branch, it mostly runs, and so the team keeps going. The author compares it to buying a luxury car on a credit card — the debt is not visible, only the car. Problems surface later as a bug nobody can fix, because nobody can answer where the data comes from. One developer, he writes, replies with a link to an AI conversation; asked which part to read, the answer is “probably all of it”.

Everyone in the story is failing

Herrengt is explicit that the blame is distributed. The engineer opening a 25,000-line pull request should have stopped the agent far earlier, understood the change and split it into reviewable pieces; the reviewer should have refused it; whoever added Kafka or serverless pieces to the stack should have been able to justify them. Reverting a decision is far harder than making it: adding tables and columns takes an AI minutes, while removing them means migration plans, paying customers, and the risk of orphaned foreign keys — all while five more changes get merged.

The widening salary gap

The author's conclusion is not that technical debt is new; it is that the limit on how fast bad decisions accumulate has disappeared. If implementation is cheap, judgement is what companies pay for, and he expects AI to push salaries further apart: good engineers become more valuable because fewer people around them are needed for implementation work, while weaker ones become expensive to hire. He argues the same dynamic will extend beyond software to most knowledge work — “at some point, someone still has to know what is going on”.