AI agents are erasing software engineering's middle class, essay argues
The essay opens with a before-and-after scene. In 2020, a senior engineer in charge of code quality goes on holiday, comes back to find the codebase a mess (new tables added to denormalise the database, Kafka or serverless bolted on without justification) and is still able to fix it. In a hypothetical 2026, the same engineer has not even left for the weekend: on an ordinary Monday morning there are already 7 pull requests waiting, the first one alone at +24,506/-3,938 lines with an AI-generated description, and the team has produced more change since Friday than it used to produce in weeks. The author calls this 'AI removed the speed limit': agents let people skip the conversation about how to build something and go straight to a working-looking pull request, so weak engineering cultures fail much faster than before.
The essay continues the scene to illustrate the cost. A recurring bug resists a fourth fix attempt; the engineer who built the feature does not know where the underlying data comes from and has to ask an AI assistant, at one point saying not even 'Fable' can figure it out. Two engineers sit reading an AI chat log neither of them can verify. When a colleague asks why a piece of code exists, the answer is a link to a Claude conversation containing roughly 15 rounds of back-and-forth, apologies and changed minds, somewhere in which the actual design decision is supposedly buried. By the time 13 PRs are stacked up, a coworker says they have run out of daily 'Fable' usage and will run a fix loop the next day instead.
The author draws a list of failures from the scene: the engineer who opened a 25,000-line PR should have stopped the agent long before reaching that size, broken the work into smaller pieces and questioned each new abstraction; the reviewer should have refused to review something that large; whoever added Kafka should have been able to justify it; whoever built the feature should have been able to explain where its data comes from without sending a chat log. The essay is careful to say technical debt is not inherently bad, only that you have to know you're taking a shortcut. Reversing a bad decision, it argues, is much harder than making one: an LLM might add a set of database tables and columns in 10 minutes, but removing them safely means a migration plan, avoiding downtime for paying users, and guarding against orphaned foreign keys. Meanwhile a single person can generate 20,000 lines of code in an afternoon, so while one bad decision is being untangled, the essay says five more get merged.
From there the author moves to the labour-market argument. Bad engineers were always a liability, but there used to be a limit to how fast that liability could compound; now implementation is cheap and, in the author's view, companies pay six-figure salaries in cities like London or San Francisco for good decisions, not for turning a specification into code, which is why firms claiming 'software is solved' still pay top salaries to attract top people. The author's stated bet is that AI pushes salaries further apart, with the bar for employability set by whatever the current best model can do: good engineers become more valuable because they need fewer people around them for implementation, while bad engineers become more expensive to hire or, per the author's earlier writing referenced here, see the 'vibe coder' career path close off entirely. The essay ends by extending the claim beyond software: the author believes AI will make the best people in most knowledge work far more productive and the worst almost impossible to hire, because mistakes can now be made faster than anyone around them can realistically review or catch.
Key facts
- In the essay's hypothetical 2026 scene, a single PR runs +24,506/-3,938 lines with an AI-generated description, and the reviewer's queue grows from 7 PRs to 13 within one Monday.
- The essay estimates an LLM can add a set of database tables and columns in 10 minutes, but says undoing that shortcut requires a full migration plan to avoid downtime and orphaned foreign keys.
- A person can generate 20,000 lines of code in an afternoon, per the essay, so while one bad decision is being untangled, five more get merged in the meantime.
- Central claim: AI removed the previous speed limit on bad engineering decisions, so projects with weak engineering culture now fail much faster than before agents existed.
- The author's bet is that AI widens pay gaps between engineers: good engineers become more valuable because they need fewer people for implementation, while bad engineers become more expensive to hire, and the author expects the same split to spread across most knowledge work.
Why it matters
The essay's core argument is that AI coding agents did not just speed up good engineering, they removed the natural brake that used to slow down bad engineering. When implementation was expensive, weak decisions accumulated slowly enough that someone usually caught them. When an agent can produce 20,000 lines in an afternoon or add database tables in 10 minutes, that brake is gone, and the author argues this changes what a team's engineering culture can survive.
Who it affects
The essay is written from the perspective of a senior engineer or tech lead responsible for code quality and review, and its scenario centers on engineering teams with agent-assisted workflows. The author explicitly extends the argument beyond software to 'most knowledge work' in general, though that extension is stated as personal belief rather than something demonstrated in the piece.
How to use it
The essay's own scenario doubles as practical advice: treat an oversized AI-generated PR, like its own example of +24,506/-3,938 lines, as a reason to stop rather than something to wave through review. It argues reviewers should refuse to review changes that large, that whoever introduces a new piece of infrastructure should be able to justify it without an agent's help, and that whoever built a feature should be able to explain it without linking a chat transcript. Technical debt is treated as acceptable only when the team knows it is a shortcut, not when it appears as an unexamined side effect of agent output.
How solid is it
This is a personal blog essay, not a study or reported piece. Its central 2020-versus-2026 scene is an admitted hypothetical used to illustrate an argument, not a real reported incident; it names real tools like Kafka, Claude, and Fable, but no company or individual other than the author is named. The claim that AI widens the pay and skill gap between engineers rests entirely on the author's own stated bet and belief, with no survey, dataset, or cited research behind it.
Risks and caveats
Because the narrative scenes are illustrative fiction rather than documented events, readers should not treat the specific numbers (PR sizes, review-queue counts, rounds of AI back-and-forth) as measurements from any real team. The salary-divergence prediction is explicitly speculative, with no timeframe or magnitude given for when or how much salaries would actually diverge.
“The people who don't will become much cheaper to hire or get replaced entirely while the money gets funnelled towards an increasingly smaller number of people who can actually be trusted.”
— the essay's author