If you're offloading every customer touch to an agent before you've personally felt the friction, you're automating your ignorance. Graham's argument is more urgent now, not less.
Paul Graham's essay is essentially a permission slip to do embarrassing amounts of hand-holding before your systems are ready. Recruit users by knocking on doors, write their onboarding emails yourself, manually do whatever your product is supposed to do automatically. The point is not efficiency — it is intelligence-gathering disguised as customer service. The founders who win early are the ones who care too visibly, too specifically, for too few people.
The trap in an AI-native studio is reaching for an agent the moment something feels repetitive. But repetitive friction is exactly the signal you need to sit inside before you automate it away. Build the agent only after you have done the task yourself enough times to know where it breaks, what edge cases embarrass it, and which moments a user actually needs a human response rather than a fast one. Graham's unscalable phase is now your agent-design research phase — skip it and your agent will be confidently, efficiently wrong.
- Do the manual version yourself before you spec the agent
- treat early users as informants, not just customers
- automate the pattern only after you can describe exactly what good looks like.
