The MVP Was Always About Learning, Not the Product
Ries defined a minimum viable product as the smallest thing that produces validated learning, not the smallest thing you could ship. Founders misread it as permission to ship junk. The actual discipline is deciding in advance what you are trying to learn, then designing the experiment to answer that question cleanly. In an AI-native company, the first deployment of an agent workflow is this kind of experiment. The question is never "can the agent do this?" The question is "does this loop produce outputs we trust enough to act on?" That distinction changes what you build and what you measure.
Pivot or Persevere Requires Data You Designed For
The pivot-or-persevere meeting in The Lean Startup is only useful if the metrics on the table are real. Ries was ruthless about vanity metrics — numbers that go up and feel good but do not correspond to learning. In agentic systems, vanity metrics are everywhere: tokens processed, tasks completed, speed gains claimed. The metrics that matter are eval results, error rates caught before consequence, and the frequency with which human reviewers overturn agent decisions. Those numbers tell you whether the loop is actually tightening or just spinning.
Humans Own the Judgment the Loop Cannot Automate
Ries argued that the hardest decisions in a startup are not technical. They are about what to believe when evidence is ambiguous. In an AI-native company, the operating loop runs continuously — task, action, eval, review, learning — but every pass produces data that still requires a human to interpret at the boundary where one strategy ends and another begins. Agents can tighten the loop; they cannot own accountability for what the loop is optimizing toward. The founder's job is to keep the learning honest, define what counts as a meaningful result, and make the call when the data genuinely could go either way.
