Garry Tan, President of Y Combinator, stated that Paul Graham's oldest piece of advice has gained traction: "Do things that cannot be scaled" still holds true, as agents now allow founders to scale such tasks for the first time. The advice remains unchanged, but the physical conditions for startups have evolved.

Graham's original intent in his 2013 paper was that early-stage work must be hands-on: personally visiting clients, building websites by hand, and manually acquiring the first ten customers, as a 10% weekly growth on a small base only equals finding one more person. He himself built stores for clients during his time at Viaweb, which was sold to Yahoo for about $49 million three years later. Tan has rewritten the second principle as still governing everything, with the new variable being that agents have become the replicators of that manual work.

He observed within YC that about a quarter of the companies in the Winter 2025 batch have 95% of their codebase generated by AI, and these companies have utilized agents in roles beyond coding. This batch is expected to become one of the fastest-growing and most profitable in history. The Emergent company from the Summer 2024 batch went from public launch to nine-digit revenue in about eight months, with an annualized revenue of approximately $15 million and around 15 employees. He referred to the quality of the model as rented, while the brain is one's own; the same set of weights and the same window can yield results ranging from 2x to 100x, with the difference lying in how the work is connected.

He claims that his output is about 400 times greater than in 2013 when writing code, with a lower bound of about 8 times and a median of about 80 times after accounting for variances. He stated that the multiplier is not limited to coding; it applies equally to design, product, and growth. Skills are documented as repeatable employee tasks, and a process that is done correctly once becomes an asset. High-capacity agents costing around $50,000 to $100,000 annually are described by him as "living in 2028."

In terms of market mechanisms, what is bought is the leverage of a single founder, while what is sold is the old organization spread across headcount. Funds flow from token bills into replicable skill files rather than into staffing. The beneficiaries are small teams that can encode manual services into agents; the pressured parties are companies that still rely on adding personnel to replicate "non-scalable" actions. If the multiplier only exists in demonstrations, payrolls will not decrease; if it exists in revenue, staffing density will become the new valuation denominator.

Additionally, Graham's other advice to "create what people want" is still considered valid by Tan. He does not claim to prove that AI-generated code has led to batch growth, but states that the fastest-growing founders treat the model as labor rather than as auto-completion.

Source: Public Information