AI got dramatically better at writing code. Engineering employment kept growing anyway. This is the argument for why those two facts are perfectly consistent.
Through 2025 and into 2026, a parade of CEOs blamed layoffs on AI. The authors call this AI washingUsing AI as the convenient public narrative for layoffs that were actually driven by financial pressure, over-hiring, or restructuring.: the financials tell a different story than the press release. Click each card to decode what the reporting actually found.
It isn't just anecdotes — it's a gap visible in the survey data. Executives say AI far more often than they act on it:
If AI really made workers productive, mass firing would be self-defeating. As the authors put it: firing workers "results in the loss of precisely the tacit knowledgeThe undocumented, hard-to-transfer understanding a team holds about its own systems, customers, and history — exactly what you need to operate AI tools effectively. and organizational capital that allows workers to operate AI effectively."
Here's the core model. Software work has three layers. AI got great at the middle one — and only the middle one.
Drag the slider to crank up AI's coding ability and watch what happens to the job.
Writing code was never the bottleneck — a 2019 Microsoft study found developers spend only 9–61% of their time coding. AI compresses the slice that was never the constraint. And as the authors note: "Once a decision can be delegated to AI, it is no longer a source of competitive advantage, and the value of human decision-making migrates upward."
This echoes Fred Brooks's No Silver Bullet: AI attacks accidental complexity (the grind of expressing a solution in code) but not essential complexity (figuring out what the system must actually do). The Decide and Deliver layers are where essential complexity lives.
If code-writing were the constraint, more AI code would mean proportionally more software shipped. It doesn't. The "Writing Code vs. Shipping Code" study tracked 100,000 GitHub developers: AI agents generated 8× more lines of code — but produced only ~30% more releases. The Deliver layer is the wall.
Push the firehose to 8×. The blue bar races to the top; the green bar crawls. All that extra code piles up against testing, review, integration, and accountability — work that didn't get any faster. More output, almost the same throughput.
Much of the "engineers are obsolete" discourse quietly assumes vibe codingYou tell an agent what to do, don't supervise it, don't review the code, and only notice problems when something visibly breaks. scales to production. The authors argue it doesn't — and that the real practice is agentic engineeringA human stays in control and accountable: supervising the agent, reviewing its code, and evaluating the output against what's actually needed., which still needs an engineer. Toggle between them:
The data says these are genuinely different animals — not two points on one spectrum:
You cannot ship production software by hiring unqualified vibe coders instead of engineers. Even Simon Willison — a champion of coding agents — reports being "mentally exhausted by 11am from supervising agents." Supervision is the job; it didn't disappear, it intensified.
The final pillar is economics. Software is highly price-elastic: when it gets cheaper to build, we don't build the same amount for less — we build vastly more. This is Jevons' paradoxWhen a resource becomes more efficient/cheaper to use, total consumption can rise rather than fall, because cheapness unlocks demand that didn't exist before.. Drag the cost down and watch total demand respond:
Automation shrank farm jobs because we can only eat so many calories — demand is capped. Software has no such ceiling: demand has "grown a millionfold," and a modern car alone runs on roughly a hundred million lines of code. The authors: "If there is a ceiling to the demand for code, we are nowhere near it."
FORTRAN, COBOL, and SQL were each sold as the tool that would let non-programmers write software and make engineers unnecessary. Each instead raised the ceiling and created more engineering work. The democratization story is old; the headcount kept climbing.
One claim trips people up most. The authors agree AI is reshaping the field — so which of these is their actual position?
AI compressed the layer of software work that was never the bottleneck. The bottleneck — deciding what to build and being accountable for delivering it — is still human, and cheaper code only raises how much of both the world demands.
The value didn't vanish when the middle of the sandwich got automated. It migrated up and out — toward judgment, specification, and ownership. "Software engineer" grows where "programmer" once was. Watch the structural shifts, not the headlines.
The authors' honest caveats, worth keeping: a Federal Reserve study shows a ~3-point annual slowdown in engineer growth post-ChatGPT (though it can't see self-employment); some indirect losses are real (Chegg, Stack Overflow); and whether poorly-supervised AI use is rising in high-stakes work is the open question to monitor.