If you're handing work to agents before you've done that work yourself, manually and awkwardly, you don't yet know what the agents should optimize for. That's the problem Paul Graham's essay quietly solves.
Do Things That Don't Scale, Paul Graham's essay from 2013, makes a case that the scrappy, embarrassing, one-at-a-time work founders avoid because it won't last is exactly the work that produces the insight and loyalty that does last. Recruiting users by hand, personally fixing their problems, delivering a product experience that only works because you're backstage frantically pulling levers — none of it survives growth, and that's fine. It teaches you the real shape of the problem before you've locked in a solution.
The temptation with AI-native tooling is to automate before you understand, which is just the old premature scaling mistake wearing a GPU. If you've never done the outreach yourself, never handled the edge case at midnight, never watched a user squint at the output your agent produced, you can't write the evals, the guardrails, or the fallback logic that makes the agent trustworthy at scale. Graham's essay is essentially an argument for earned abstraction — you get to hand the work to agents only after the unscalable phase has told you precisely what quality looks like.
- Do the manual version first so you know what right looks like before you define it for an agent
- the users you recruit by hand become the trust anchors who tell you when automation breaks
- designing human judgment back into an AI-native product is easier when you've personally held that judgment at least once.
