How AI coding agents will reshape software engineering and knowledge work

Executive overview

AI coding agents have crossed a threshold where non-engineers can ship production software at scale, and professional engineers using them are an order of magnitude more productive. The combine harvester analogy applies: food production rose dramatically while farmer headcount collapsed. Demand for software may rise 10–100x, but AI will meet that demand — not humans.

The job of "software engineer" as it exists today will not survive the decade; what survives is high-agency problem-identification and taste.

The capability shift is already underway

• Tom Blomfield built a 35,000-line production app with voice agent using only prompts — zero lines written manually
• After ~5,000 lines he stopped reading the code; prompt, auto-accept, iterate
• 90 minutes on a train: new blog software, hosting set up, 15 years of posts migrated
• At YC: ~0% of companies used AI for most code two batches ago; 25% last batch; 33–50% now
• Tools (Lovable, Cursor, Windsurf, Claude Code) improved sharply in the last 6–9 months

Why "it's not good enough yet" is the wrong argument

• Current tools are imperfect but have a clear improvement trajectory over 3, 6, and 12 months
• Christensen's innovators dilemma: disruptive tools always look like toys until they don't
• Arguing they'll never be good enough requires believing improvement stops now
• Jevons paradox (more efficiency → more demand) is real, but AI will fulfil most of that demand, not humans
• The ratio of output per human will exceed even large demand increases

What remains human

• Taste and obsession — the best software has one person behind it who is obsessed with quality
• Problem identification: understanding what real human problems are worth solving
• No current methodology makes AI reliably obsessed with solving a specific problem
• Physical work (surgeons, plumbers, electricians) is structurally protected
• Regulatory gatekeeping will slow displacement in law and medicine — not stop it

Knowledge work beyond coding

• Legal: Legora (YC) succeeded despite the received wisdom that "lawyers never buy software"
• Pressure is now top-down: boards ask every CEO "what is AI doing to our business?"
• Using AI is becoming table stakes — like email; not using it will be a competitive disadvantage
• Cost of knowledge work falls → massive consumer surplus, but roles must change
• Transition period could be severe: hundreds of millions displaced faster than retraining is possible

Advice for founders and people entering the workforce

• Stay current with the latest tools — early fluency compounds into a durable career advantage
• Over-index on understanding people and identifying real problems; building is becoming cheap
• Smaller engineering teams (2–4 people) will build what previously required 40
• Fewer team boundaries → cleaner ownership → better product design and user experience
• Industries newly open to software (law, education, medicine) represent the biggest greenfield opportunity
• Right now is the best time in history to start a company from scratch