Eric Ries wrote The Lean Startup for a world where building was expensive; agents just made building nearly free, which makes his central warning — that output is not progress — more dangerous to ignore, not less.
The book's origin story is a well-run company executing flawlessly toward a product nobody wanted. Ries's fix was to treat a startup as an experiment under extreme uncertainty: name your leap-of-faith assumptions, build the smallest thing that tests one, and measure whether real customer behavior changes. Progress is validated learning — what you now know that you didn't — never features shipped or dashboards that only go up. And the pivot-or-persevere call gets a scheduled meeting, so the hardest decision becomes a discipline instead of a mood you avoid.
Agents collapse the build phase of build-measure-learn from months to hours, which means the loop now stalls entirely at measure and learn. That's a trap dressed as a gift: an agent can produce a quarter's roadmap in a weekend, all of it untested, the most seductive vanity metric ever invented. The AI-native move is to point agents at experiments — instrumentation, cohort splits, the smallest testable slice — while the founder keeps the one job that doesn't delegate: reading the evidence and deciding whether to pivot or persevere.
- velocity from agents is vanity output until a customer behavior changes
- spend agent capacity on experiments and measurement, not just features
- keep pivot-or-persevere a scheduled human decision, because that judgment is now your whole job.
