The Sandwich Model: Why AI Can't Replace Software Engineers

The article refutes the narrative of AI-driven mass layoffs in software engineering, labeling many such claims as 'AI washing' for financial restructuring. It posits that while AI can automate code execution, humans remain essential for the 'decide' and 'deliver' phases of the development sandwich. Ultimately, the authors predict that AI will increase software demand and complexity, keeping skilled engineers at the center of the profession.

Key Points

• Many reported AI-driven layoffs are 'AI washing' used by companies to mask financial struggles or pandemic-era over-hiring.
• Software development is a 'decide-execute-deliver sandwich' where AI only automates the middle execution layer, leaving the ends to humans.
• Writing code was never the primary bottleneck in software engineering; the real challenges are requirements specification and accountability.
• There is a critical distinction between 'vibe coding' (unsupervised and risky) and 'agentic engineering' (human-in-the-loop and professional).
• Economic principles like Jevons' paradox suggest that cheaper code production will lead to higher demand for software and engineers.

Sentiment

The community is mixed and skeptical, with meaningful agreement that AI does not currently replace the full accountability-bearing role of software engineers, but strong resistance to the article's confident long-term framing. Hacker News largely accepts the article's distinction between code generation and delivered software, while remaining unconvinced that this distinction fully protects jobs or team sizes.

In Agreement

• Software engineering has repeatedly automated parts of itself, and higher productivity has historically expanded expectations and demand rather than eliminating the profession.
• The hard parts of useful software are deciding what should exist, integrating it into messy contexts, verifying behavior, maintaining it, and being accountable for consequences.
• Current AI works best when an expert guides, reviews, tests, and constrains it; unsupervised generation is not the same as professional delivery.
• Liability, trust, and institutional responsibility keep humans in the loop because model vendors and autonomous agents do not bear real accountability for failures.
• Narrow implementation work may be automated, but adaptable engineers who understand systems, users, and operations remain valuable.

Opposed

• The article overstates the permanence of the decide and deliver layers, because improved AI capabilities and tooling could automate more of those responsibilities over time.
• Even if some engineers remain necessary, teams may need far fewer people once code production and routine engineering tasks become much cheaper.
• The claim that AI will not replace engineers sounds like professional wishful thinking, especially in a labor market where management may use AI narratives to justify cuts.
• Personalized AI-generated apps could reduce demand for many small commercial products and routine developer services, especially for low-stakes software.
• Code generation was never a trivial bottleneck, so compressing it can materially change staffing, timelines, and the value of certain engineering specialties.