The Future of Software Engineering is SRE
Swizec's argument that as AI makes writing code trivially easy, the scarce skill becomes keeping software running reliably. The demo is the easy part. The other 190% -- uptime, incident response, failure recovery, legal accountability -- is what separates toys from services.
Key Quotes
"The first 90% to get a working demo is easy. It's the other 190% that matters."
"People don't buy software, they hire a service."
Key Themes
The article uses a cautionary tale: an accountant who builds a spreadsheet automation that saves hours weekly, then spends every vacation debugging it. Richard Feynman's "computer disease" -- people enjoy tinkering with automation but dread the responsibility of running it at scale, indefinitely.
The operational demands are concrete: uptime guarantees, defect identification speed, proactive issue detection, secure data handling, timezone-aware incident response, failure recovery without data loss. And the killer question: "Will you sign a legally binding guarantee?"
This reframes the entire AI-coding revolution. If agents can write code 10x faster, that just means we produce 10x more code that needs operating. The bottleneck moves from creation to sustenance. Organizations that master operations win; those that only master generation produce expensive liabilities.
The connection to Building Production-Ready Voice Agents is direct -- Gulati's team spent 50% of effort on admin portal and observability, not the voice agent. That's SRE in practice. Designing a Passively Safe API shows what SRE-aware API design looks like. And Write Only Code pushes the question further: if nobody reads the code, your operational signals become the only feedback loop.
Critical Analysis
The argument is sound but the article pulls its punches. It establishes the principle well -- operations > generation -- but doesn't go deep into what SRE looks like in an AI-native world. How do you do incident response when nobody wrote the code? How do you debug systems where the implementation was generated by three different models across six months?
AI Handles Incidents, Engineers Lose Touch with Their Systems answers one piece of that: the AI SRE can now run the response, but every routine incident it auto-resolves is practice the human responder didn't get. The proposed fix is incident simulation — rehearsing failure on a schedule, the way pilots do, rather than letting a real outage be the training ground.
The Feynman reference is apt. The accountant parable is effective. But the piece would benefit from concrete examples of AI-generated systems that failed in production, not just hypothetical warnings.
Sources: The Future of Software Engineering is SRE Last updated: 2026-05-14