Risk Enumeration Comes Before Headcount
Most founding teams hire by instinct — someone impressive is available, so they join. Khosla's method inverts that completely. You write the risk register first: what are the five things most likely to kill this company? Only then do you ask who has faced that exact problem and come out the other side. For an AI-native company, those risks almost always include evaluation failure, trust collapse in production, and the organizational gap between the people who set goals and the agents executing them.
The Agent Stack Is Its Own Risk Surface
Building with agents means your risk register needs entries that did not exist five years ago. Workflow design that degrades silently, eval harnesses that never get built because the team shipped without them, tools scoped so loosely they return noise — each is a named, hirable risk. A founder applying gene pool engineering asks not just who has shipped great software, but who has built reliable agentic systems where the failure modes were caught before customers felt them, and who designed the feedback loops that kept them caught.
Judgment Is the Risk No Agent Retires
Agents can execute a workflow, choose a path through branching logic, and return structured results with measurable quality. What they cannot do is own accountability, set the goals in the first place, or make the tradeoffs that require someone to be answerable when things go wrong. That is the risk that sits at the top of every AI-native risk register, and the hire who retires it is someone with a demonstrated history of holding judgment under pressure — not someone who simply understands the technology.
