Start With What the Agent Gets Wrong
Graham's central move is to look for things that are broken in your own life. For an AI-native founder, that search has a specific texture: watch where your own agentic workflows quietly fail, where someone manually fixes the output before it ships, or where the feedback loop produces confident nonsense. The operating loop — task, tools, context, action, eval, human review, learning — generates real problems at every handoff. Those handoffs are the ideas.
The Toy That Agents Make Possible
Graham notes that the best early ideas look like toys — too narrow, too weird, dismissed by the mainstream. In an AI-native company, the toy version is often an agent handling one tedious workflow for a small group that desperately needs it automated. The loop is embarrassingly thin. The eval criteria are hand-coded. That smallness is not a flaw; it is the proof of concept, and every pass through the loop leaves data that sharpens the next run.
Judgment Is the Moat, Not the Model
Graham warns against building what sounds impressive rather than what someone urgently wants. That warning maps precisely onto AI-native product design: the model is not the company. What makes the thing defensible is the human judgment layer — the goals set, the boundaries defined, the accountability owned — wrapped around reliable, well-scoped agentic work. Distribution, trust, and taste become scarce once software gets cheap to create, so founders who notice real problems and design trust in from the start hold ground that pure automation cannot.
