If your first hundred users don't feel personally obsessed over, you won't know what to automate — and right now, agents let you do that obsessing at a cost that has no precedent.
Paul Graham's essay is a permission slip most founders misread. The argument isn't that you should stay small — it's that the unscalable phase is epistemically necessary. Going door to door, onboarding users by hand, doing customer support that looks like concierge service: these aren't embarrassing stopgaps. They are the mechanism by which you learn which problems are real, which users will evangelize, and what a working product even feels like from the outside. You cannot design that knowledge in advance.
The classic tension in Graham's argument was time: founders have a finite number of hours to spend being ridiculous attentive. Agents dissolve that constraint without dissolving the discipline behind it. You can now send a personalized onboarding sequence to every new user, drafted by an agent that knows their use case, reviewed by you before it sends. You can run a manual-feeling support loop — agent drafts, human approves, user feels heard — at fifty users or five hundred. The trap is using that leverage to skip the learning phase rather than extend it. The point of unscalable work was never the labor; it was the proximity to what is actually happening.
- Stay in the unscalable phase longer than feels comfortable, because agents now let you afford it
- use agents to do the handcraft, not to automate past the moment of learning
- the founder who approves every agent output in week one is doing fieldwork, not busywork.
