← Wiki

Vibe Coding and the New Software Labor

ai-codingfuture-of-workai-agentsproductivity

The Term and Its Implications

“Vibe coding” entered the lexicon to describe a new mode of software creation: directing AI to write code through natural language, iterating by feel rather than by precise technical specification. The term captures both the promise (anyone can build software) and the skepticism (is this really engineering?).

Andrej Karpathy, who coined the term, reportedly “doesn’t call himself a coder anymore” — which Nityesh finds telling: “the guy who invented the term is saying that he doesn’t even want to be identified as a programmer.”

The Democratization Thesis

Vibe coding dramatically lowers the barrier to software creation. Jason Lemkin, after building B2B products with AI coding tools, reports that the approach works for certain categories of applications. Santiago Valdarrama captures the confidence it creates: a person who “can’t tell Python from C++ — has never written a single line of code — yet feels he can build anything he wants.”

This confidence is simultaneously the promise and the risk. The promise: orders of magnitude more people can now prototype, ship, and iterate on software. The risk: “The most fascinating aspect of vibe-coding is how it has convinced so many people to believe they are better and more capable than they really are.”

Unevenly Distributed Productivity

Jeffrey Wang maps the productivity distribution across experience levels:

The gains are not uniform. Experienced engineers use AI coding as leverage for the work they already understand. Novices use it as a substitute for understanding they don’t have. The gap between these two modes determines whether the output is robust or fragile.

Alex Lieberman clarifies the misconception: “A huge misconception non-technical folks have is that engineers & CEOs vibe coding is the same as their vibe coding.” The same tool in experienced hands produces structurally different output than in inexperienced hands.

The Technical Debt Question

Dharmesh Shah (HubSpot CTO) frames the central open question: “How do you make sure that the compounding value you are getting from the use of agentic coding exceeds the interest on the technical debt that AI creates along the way?”

AI-generated code ships faster but may accumulate structural problems that compound over time. The hope is that AI itself can eventually pay down this debt as models improve — but that remains unproven.

Levie extends this to the broader meaning of code quality: “The code may not be the thing you ultimately care about or are even responsible for. The output of what the code creates — the end product, the experience, etc. — is what matters.” If AI can maintain code quality through automation, the human’s job shifts entirely to defining outcomes.

The Emerging Workforce Pattern

Cody Schneider argues that “the most effective employees will be the ones that orchestrate 100+ AI agents across hundreds of different tools” — a vision of work where coding skill matters less than orchestration skill. The new software labor is not writing code but directing AI systems that write it.

Omar Khattab pushes back from the research side: “I will be so disappointed if the way we build software with AI remains vibe coding.” He argues for more structured approaches — suggesting the field will split between casual vibe coding for simple applications and rigorous AI-assisted engineering for complex systems.

Aakash Gupta reports that the Browser Company’s founder warns: “If you don’t work Claude Code-native ASAP your startup is dead” — framing AI-native development not as an option but as a competitive requirement.

Connection to the Broader Thesis

Vibe coding is a case study in Enabling vs Replacing Technologies:

  • For experienced engineers: enabling. They ship more, tackle harder problems, and focus on architecture rather than syntax.
  • For non-engineers entering the field: enabling in the short term, potentially replacing in the long term if they never build deep understanding.
  • For organizations: it shifts the Bounty and Spread dynamic. More total software gets created (bounty), but the quality gap between AI-leveraged experts and AI-dependent novices widens (spread).