If you're handing agents real work and watching them fumble toward something useful, you're not doing it wrong — you're doing it exactly as the founders in Livingston's book did it.
Founders at Work, Jessica Livingston's collection of interviews with people who built companies now considered inevitable, delivers one persistent finding: none of them knew what they were building when they started. The pivots weren't failures dressed up in retrospect — they were the mechanism. Almost every founder describes a version of the product that would embarrass them now, combined with a stubbornness about staying in the room until something worked. Grand strategy was largely retrofitted after the fact.
The AI-native equivalent of Livingston's embarrassingly small beginnings is the embarrassingly narrow agent task you deploy on Monday. Founders who treat their first agent workflow as a prototype to be observed rather than a product to be defended are doing exactly what her interview subjects describe — staying close, staying humble, and letting the actual behavior of the system rewrite the spec. The judgment you hold as a human founder is not about vision; it is about noticing which fumble contains the real signal.
- Deploy the narrow thing and watch it closely before scaling the scope
- treat every agent failure as a pivot signal, not a deployment failure
- the founder's job is pattern recognition on messy output, not clean orchestration from above.
