Vibe coding: how AI is reshaping software engineering roles

Executive overview

Most software is now being written by LLMs. A quarter of founders in the current YC batch report that over 95% of their codebase is AI-generated. The role of the engineer is splitting into two distinct tracks: product thinker and systems architect.

Vibe coding accelerates zero-to-one dramatically. But scaling from one to many users still requires classically trained systems thinkers — and no current tool handles that well.

The core insight: taste and systems thinking are now more valuable than the ability to write code quickly.

What founders are actually doing

• Cursor is the dominant IDE; Windsurf gaining ground by auto-indexing the full codebase
• Claude Sonnet 3.5 remains the leading code-gen model; reasoning models (o1, o3) now competitive
• DeepSeek R1 emerging as a viable alternative; Gemini used mainly for its long context window to one-shot bug fixes
• Devin used only for small, isolated tasks — doesn't understand the codebase well enough for serious features
• Some founders with sensitive IP are self-hosting models
• Parallel prompting: running two Cursor windows on two features simultaneously

How workflows have changed

• Code is treated as disposable: reroll instead of debug when rewriting is cheaper than fixing
• Debugging remains the hardest unsolved problem — current models need explicit, spoon-fed instructions
• Reasoning models (o3) are meaningfully better at debugging than earlier generation models
• Founders describe themselves as product people and reviewers, not engineers
• Decisions to scrap and rewrite are less emotionally loaded when code costs nothing to reproduce

The two emerging engineering roles

• Product engineer: high taste, talks to users, translates problems into working software via AI
• Systems architect: deep CS fundamentals, designs infrastructure that scales, can audit what AI produces
• The zero-to-one phase suits vibe coding; the one-to-N phase still requires systems thinkers
• Historical parallel: Rails and PHP enabled fast shipping but required rewrites at scale (Twitter fail whale, Facebook's HipHop compiler)

What still requires humans

• Debugging: reading code, tracing logic errors, understanding what the system is actually doing
• Taste: recognising when AI output is bad — you can't catch it without enough training to judge
• Calling out bullshit: technically weak founders get misled by engineers and, increasingly, by AI agents
• Systems design at scale: low-level work that current tools explicitly cannot do

The AI-native engineering generation

• Some current YC founders have never coded without AI tools — they learned to program in the cursor era
• Math and physics backgrounds give the systems thinking needed to supervise AI output without classical CS training
• Good enough engineers will be abundant; exceptional engineers will still require deliberate practice
• Picasso analogy: abstract mastery comes from classical training first, not instead of it

Implications for hiring and screening

• Current engineering hiring has not caught up — whiteboard interviews are increasingly irrelevant
• Stripe and Gusto moved early toward productivity-based screening; that approach is now the baseline
• New screens should test: debugging ability, code review, systems design, and taste
• The question of whether to allow LLMs during interviews is unresolved — but old questions are now trivially solved by AI
• Key screen: can the candidate identify when AI output is wrong?