Map Your Risks Before You Map Your Org
AI-native companies carry a particular class of risks that founders misname. They call them "technical challenges" when they are actually trust failures waiting to happen — agents that drift off-task, workflows that look agentic but have no real branching, evals that get built after things break in production. Khosla's method forces you to name these honestly and specifically before you hire anyone. The discipline is the same: write down your five biggest points of failure, then ask who has retired each one before, because vague hiring produces vague teams that produce vague systems.
The Eval Risk Is Your First Hiring Problem
Reliable agentic work requires evaluation harnesses built before the agent ships, not after. That is a specific skill with a specific track record, and almost no founding team has it. Gene pool engineering applied here means your first critical hire is not the most impressive generalist — it is someone who has designed and run evals under production pressure. If you cannot measure agent behavior, you cannot trust it, and if you cannot trust it, your product is an expensive demo. That risk has a face; go find the person whose face it already fears.
Judgment Gaps Are Risks Too
Human judgment — knowing when to escalate, when to override, when to hold the line on accountability — is not a soft virtue in an autonomous system, it is a hard architectural requirement. If no one on your team has ever designed the boundary between agent autonomy and human override, that gap is a risk on your list, not a detail to figure out later. Khosla's framework insists you treat it as a first-class engineering problem and hire accordingly. The founding team should be able to point to a specific person and say: they have drawn that line before and it held.
Taste and Orchestration Are Now Hiring Criteria
As software becomes language-directed, the work of founding is increasingly orchestration — deciding what agents do, what tools they touch, what context they receive, and what output quality actually looks like. That is a taste and judgment function, not just a coding function. Gene pool engineering in this environment means you need people who have shipped opinionated systems under ambiguity, not just people who can prompt cleverly. The risk is mediocrity laundered through automation. The hire that retires it is someone who has held a bar and enforced it when the machine wanted to lower it.
