Noticing Is Still Human Work
Graham's core move is to stop brainstorming and start observing your own friction. That logic holds exactly when Software 3.0 collapses prototyping to hours. The bottleneck was never generating code — it was knowing what was worth building. A founder directing agents in natural language faces the same trap Graham describes: the scarce skill is recognizing a real problem, and that recognition still lives entirely with a human who has lived inside the domain.
Small and Desperate Beats Big and Vague
Graham insists on starting with a small group that urgently needs the thing, not a large group that might want it someday. For an AI-native company, this is structural advice. An agent workflow built around a crisp, urgent problem can be tested, evaluated, and tightened in days. A workflow built around a fuzzy hypothesis just executes the fuzziness at scale and speed. Precision in the problem statement is what makes the harness work.
Toys at the Edge of the Newly Possible
Graham notes that breakthrough ideas often look like toys and live at the frontier of what is technically new. Today that frontier is agentic software — tasks that agents can now execute reliably that were impossible two years ago. A founder's job is to notice a real problem that sits exactly at that edge, where the capability just arrived and no incumbent has redesigned their product around it yet. That matching of genuine pain to new capability is judgment no model provides.
Distribution and Trust Do Not Compress
Graham's essay focuses on ideation, but the translation to the present requires honesty about what comes after. Finding a sharp problem and building fast with agents still leaves a founder facing the same durable constraints: who trusts you enough to try the thing, and how do you reach them. Software getting cheaper to produce makes those questions harder, not easier, because every competitor has the same acceleration. The idea is the starting point; distribution and earned trust are where the work actually compounds.
