If your agents are shipping output nobody measures, you're just burning compute on vanity. The build–measure–learn cycle Ries built for human teams gets sharper and more dangerous when the loop runs at machine speed.
Eric Ries wrote The Lean Startup as a corrective to the assumption that more output equals more progress. The real currency is validated learning — a disciplined answer to whether you changed behavior in the world, not whether you shipped something. MVPs are not rough drafts; they are experiments with hypotheses attached. The pivot is not a failure word; it is a structured response to evidence. Every metric worth tracking connects causally to something a customer actually does.
Agents collapse the build side of the cycle almost to zero, which sounds like a gift until you realize the measure and learn sides stay stubbornly human-paced. Founders are now drowning in output — agents drafting, agents calling, agents classifying — while the instrumentation to tell signal from noise lags badly behind. Ries would call this the classic growth-before-learning trap, just running faster. The discipline his framework demands is more urgent, not less: define the falsifiable assumption before the agent fires, route the result to a human judgment checkpoint, and treat every agent workflow as a hypothesis about customer behavior rather than a solved problem.
- Attach a hypothesis to every agent task before it runs, not after
- instrument agent output for causal metrics or you are measuring activity, not learning
- treat the pivot-or-persevere decision as a human checkpoint the agent cannot own.
