The hard part of building an AI-native company isn't the agents — it's knowing which problem deserves one. Graham's framework cuts through that before you waste six months automating the wrong thing.
Paul Graham's essay on startup ideas is less a how-to than a warning: the founders who sit down to "think up" ideas are already behind. The ones who win noticed something genuinely annoying in their own lives, often something so specific and unglamorous that it didn't feel like an opportunity at all. He makes much of the fact that the best ideas tend to look like toys or niche fixations, and that a small group of people who desperately need the thing is worth more than a large group who might be mildly interested someday.
The translation for an AI-native studio is almost uncomfortably direct. Agents are good at removing friction from processes that already exist and repeat, which means your idea selection problem is now more exposed than ever — you can build fast, so building the wrong thing is a faster failure. Graham's instinct to live inside a problem domain rather than brainstorm at it becomes a competitive method: the founder who notices a genuine workflow that breaks daily is sitting on the exact kind of repetitive, definable task that an agent can own. The toy-like ideas he praises are precisely the ones that look like thin automation wrappers until suddenly they aren't.
- Look for problems you hit personally and repeatedly before you consider what agents could theoretically handle
- a small desperate user group is a better signal than a broad addressable market when you're staffing a workflow with agents
- the "this barely seems worth building" feeling is often the tell that you've found something real.
