AI startups, agency, and the good timeline: a YC founder retreat debrief

Executive overview

AI is creating an entirely new category of businesses — ones that were never economically viable before. Startups selling AI agents to enterprises are seeing growth rates that would have seemed delusional two years ago: batches averaging 10% week-on-week, companies going from zero to $12M ARR in 12 months.

The driver is not just better models. It is agency and taste — knowing what to ask, how to prompt, and how to evaluate output. The companies winning are technically deep, iterate relentlessly, and treat their eval sets as their most valuable asset, not their codebase.

The real moat in AI is not access to models — it is the gold-standard labeled eval set nobody else has.

Growth rates have fundamentally reset

• YC batches averaged 10% week-on-week growth across the full cohort — previously only the top one or two companies hit that bar.
• Reaching $1M ARR in 6 months is becoming the baseline; founders are openly targeting $1M to $20M in a single year.
• Demand is pre-existing — enterprises are under internal pressure to adopt AI, so founders do not need to convince; they need to build something that works.
• Heavily technical CEOs with weak sales skills are winning large enterprise contracts because product quality is the differentiator.

Why AI agents for business are winning

• The hard part is not accessing a model — it is building software that performs work at human-equivalent quality (customer support, sales calls, etc.).
• Founders are inventing new prompting and reliability patterns because off-the-shelf attempts hallucinate and fail; persistence and wizardry around prompting make the difference.
• Usage-based pricing ties cost directly to ROI, shortening sales cycles — the product pays for itself in the same month it is deployed.
• Example: Jerry (insurance) cut customer support budget by half after deploying GPT-4, turned a $10M/year burn into a cash-flowing business compounding at 50%+ annually.

Evals and prompting as the new core asset

• Founder talks at the retreat focused heavily on evals and testing — a first for a YC conference.
• One founder: the most valuable thing his company built is the eval set, not the codebase.
• General-purpose data is cheap and plentiful; a meticulously human-labeled gold-standard eval set is rare and defensible.
• This is why the "ChatGPT wrapper" criticism misses the point — models commoditise, but prompting craft and eval sets do not.

The shift in how products get built

• Cursor went from single-digit adoption to ~80% of a YC batch in one batch cycle; Cursor hit ~$50M ARR.
• One startup's designer stopped using Figma entirely — designing in Claude via text-to-code, iterating faster than mockup workflows allowed.
• Hiring signal: founders who ask candidates whether they use AI code tools are filtering out those who say no.
• Interview standards are evolving: measure raw output in a fixed time window, not whiteboard CS theory.

Smaller teams, bigger outcomes

• The blitzscaling era (hire fast, raise huge, win via network effects) is over for this cohort.
• Companies hitting $10–20M ARR targets are doing it with fewer people and have not raised a Series A.
• The API-line metaphor: above the line means having agency; AI amplifies that leverage dramatically for solo founders and small teams.
• Stack rewrites happen every few months — the best startups throw away approaches that stop working and rebuild without sentiment.

The two forks: control vs. agency

• The bad AI path: technology used to constrain and control people.
• The good path: maximising human agency, creativity, and potential — giving non-coders the ability to build apps, non-artists the ability to create images.
• The right objective function (predict the next token) solved the alignment fear from 2015: unlike survival-driven intelligence, this AI has no drive to perpetuate itself.
• OpenAI was started as a nonprofit moonshot against Google's monopoly; 10 years later the market has at least six competing frontier models including an open-source one.

Long-run economic picture

• Google referral traffic down ~15% year-on-year; Stack Overflow down 60% since 2022 — early-adopter behaviour predicts the mainstream shift.
• AI will drive massive deflation in machine-producible goods (medical care, housing, infrastructure) while human-time activities (live performance, craftsmanship) retain distinct value.
• The dual economy: "machine money" (abundant, cheap, AI-produced) vs. "human money" (scarce, personal, human-produced).
• Wealth creation historically outpaces job displacement — 97% of people were farmers; now it is under 3%.