Why AI hasn’t replaced software engineers, and won’t

Below top-k · 15

• evidenceHeadline AI layoff stories collapse on inspectionc 0.70

  Block, Snap, and Intuit all announced cuts framed as AI-driven, but the actual drivers were activist investors, restructuring, or unrelated cost-cutting. The Snap cuts even targeted divisions that don't match an AI-displacement pattern.

• evidenceAI-washing of layoffs is widespreadc 0.70

  59% of U.S. hiring managers admit they emphasize AI when explaining cuts because it sounds better than financial constraints, and a Forrester analyst reports that nine in ten companies citing AI-driven layoffs have no mature AI app ready to do the work.

• evidenceExecutives are cutting in anticipation, not in response, to AIc 0.60

  In an HBR survey of over 1,000 executives, 21% had made large anticipatory headcount cuts versus only 2% reductions tied to actual AI implementation. The 10x gap suggests executives are themselves captured by the displacement narrative.

• contextVibe coding and agentic engineering are different practicesc 0.60

  True vibe coding means handing off without review or accountability; agentic engineering means using agents as tools while staying in control. Conflating them produces wildly wrong predictions about what AI is replacing.

• implicationProduction software cannot be shipped by unqualified vibe codersc 0.60

  Vibe coding and agentic engineering sit on a spectrum, but the practical implication for jobs is firm: companies cannot replace engineers with non-engineers running agents and still ship reliable production software.

• caveatDemocratization predictions have failed beforec 0.60

  Every major new language — FORTRAN, COBOL, SQL — was billed as letting domain experts do their own programming, and none of them actually displaced engineers. The barrier was never syntax; it's the judgment and accountability needed across the full sandwich.

• implicationThe engineer becomes a crane operator for AI agentsc 0.55

  As execution gets delegated, the human's job becomes supervising agents and keeping them in control. Most of the cognitive heavy lifting moves to the machine, but oversight becomes the bulk of the human role.

• evidenceProgrammer employment has grown from near-zero to millionsc 0.55

  Unlike agriculture, where mechanization decimated jobs because caloric demand is fixed, software demand has grown a millionfold and modern cars now contain a hundred million lines of code. The historical pattern is expansion, not collapse.

• caveatSome AI-adjacent job losses are real but indirectc 0.50

  Products like Chegg and Stack Overflow have shed workers because AI killed demand for the product itself, not because AI does those workers' jobs. The historical analogue is telegraph operators, whose work was rendered obsolete rather than directly automated.

• contextAggregate stability doesn't mean individual stabilityc 0.50

  Even if total demand for software engineering stays healthy, individual engineers will face turbulence based on firm type, geography, seniority, and adaptability. That distribution question is the subject of the next essay in the series.

• contextCompression of execution started decades agoc 0.45

  BLS started tracking programming separately from software engineering over twenty years ago, with programming work treated as lower-paid grunt work. AI is accelerating a long-existing trend of devaluing pure execution, not initiating it.

• caveatAI-vendor layoffs are reallocation, not displacementc 0.40

  When IBM or SAP cuts staff while citing AI, the honest framing is that they're shifting headcount toward fast-growing AI product lines. That's ordinary restructuring, not technology displacing workers.

• evidenceSupervising agents is exhausting, not freec 0.40

  Simon Willison reports being mentally drained by 11am from supervising coding agents. This pushes back on the assumption that oversight is cheap enough to be optimized away.

• contextCEOs are systematically prone to AI delusionc 0.40

  Aaron Levie's point is that CEOs can build flashy prototypes but never see the 90% of work needed to ship a finished product. That asymmetry helps explain why executive AI commentary so often outruns reality.

• caveatDetecting reckless AI use remains a critical data gapc 0.40

  Viral stories about agents deleting production databases go viral precisely because they're aberrant, but we currently lack the ability to measure whether poorly-supervised AI use for high-stakes tasks is actually trending upward.

Redundant with selected · 8

• claimPercent of code written by AI is a misleading metricc 0.80 · sim 0.85

  Leaders quote 65%-or-higher AI code share as if it portends mass displacement, but the metric is almost completely disconnected from labor displacement because writing code was never the bottleneck.

  overlapped with: 8x more code, only 30% more releases

• mechanismThe decide layer cannot shrink because value migrates upwardc 0.75 · sim 0.84

  Once a class of decisions can be delegated to AI, it stops being a source of competitive advantage and human judgment moves to harder decisions. Since software complexity keeps growing, there's no ceiling to this migration.

  overlapped with: The decide and deliver layers resist automation for structural, not capability, reasons

• claimLayoffs are the wrong signal of AI productivity gainsc 0.70 · sim 0.83

  Research finds AI's labor effect operates through slower hiring rather than firings. Firing existing workers destroys the tacit knowledge needed to use AI well, and natural turnover achieves the same result without the costs.

  overlapped with: AI capability gains alone won't trigger mass layoffs in software engineering

• mechanismAccountability is preserved by choice, not just capabilityc 0.70 · sim 0.83

  Even if AI becomes technically reliable, society can keep humans in the loop via liability law, sector regulation, and norms. This is a more resilient lever for controlling AI's labor impact than trying to slow capability progress.

  overlapped with: AI capability gains alone won't trigger mass layoffs in software engineering

• implicationThe sandwich pattern generalizes beyond softwarec 0.65 · sim 0.86

  Most knowledge work involves complex decisions and accountability at the ends, with execution in the middle. Ignoring this pattern is why predictions of imminent job loss in fields like radiology have repeatedly overshot.

  overlapped with: Knowledge work is a decide-execute-deliver sandwich

• evidenceSoftware engineering employment is still growing, just slowerc 0.60 · sim 0.83

  A Federal Reserve study finds post-ChatGPT software engineer employment growing about 3 percentage points per year slower than a no-AI counterfactual, but the methodology misses self-employment, which AI also makes easier.

  overlapped with: Demand for software engineers may actually rise

• evidenceDevelopers spend only a small fraction of their time writing codec 0.60 · sim 0.82

  A 2019 survey synthesis, consistent with Microsoft's own data on 6,000 developers, found coding takes 9-61% of developer time. Late-2025 blog posts from working engineers confirm that automating code generation barely moved overall productivity.

  overlapped with: 8x more code, only 30% more releases

• implicationGrowth in demand won't necessarily flow to big techc 0.40 · sim 0.83

  Most engineers already work in-house at non-software firms, and that share may grow. AI rollups buying Main Street businesses and rebuilding them AI-native could further distribute engineering work, though that trend is still speculative.

  overlapped with: Demand for software engineers may actually rise