Skip to content
El Observatorio

Foundations · 2 min de lectura

When Agents Do the Work, Graham's Failures Still Kill You

Paul Graham's "The 18 Mistakes That Kill Startups" reads like a postmortem catalog, but its underlying logic is a single claim: most startups die because they failed to make something people genuinely want, and everything else on the list is a variation of that failure. For founders building with agents this year, the list holds — but the failure modes now arrive faster and wear new clothes.

Con · estudiado y reformulado para builders AI-native“The 18 Mistakes That Kill Startups” — Paul Graham

The Bottleneck Moved, Not the Judgment

Graham's first mistake is the single founder, and his concern is not loneliness but the absence of someone to catch your errors. In a Software 3.0 environment, prototyping collapses to hours, so a solo founder can build further before anyone pushes back. That speed is a gift that doubles as a trap. The interface being natural language does not mean the thinking behind the prompts is sound. The bottleneck moved from writing code to knowing what to ask for and recognizing when the answer is good — and that judgment cannot be delegated to the agent.

Marginal Niches Are Cheaper and More Tempting Now

Graham warns against the marginal niche: a market too small to matter, chosen because it felt safe to enter. Agents make niche products trivially fast to assemble, which means founders face a new version of this failure every sprint cycle. Distribution, customer access, and trust remain the scarce things in a world of abundant software. A workflow that runs cleanly inside a product nobody wanted is still a graveyard. The question Graham pressed on — does anyone desperately need this? — is not answered by the elegance of the agentic pipeline behind it.

Platform Dependence Is an Architectural Decision Now

Graham listed derivative ideas and platform dependence as distinct mistakes, but for AI-native founders they often fuse. A product whose core behavior is entirely determined by a single model provider's API is derivative in the structural sense Graham meant: the moat belongs to someone else. Reliable agentic work comes from clear task design, good tooling, meaningful evals, and feedback loops the team owns. That architecture is also how a company builds something that cannot be trivially replicated the day the underlying model improves and everyone else gets the same capability.

Slowness Is Still Fatal, and Now It Compounds

Graham was direct that running out of money is usually a symptom of slowness — slow to ship, slow to learn, slow to quit a bad direction. An AI-native company designed around agents and continuous learning has the machinery to move faster, but the machinery requires intentional oversight. Humans still set the goals, make the tradeoffs, and own accountability for what the system produces. Founders who treat agentic speed as a substitute for clear direction are slow in the way that kills companies: lots of output, very little learning.

Prueba el Desktop OS

Explora la escena startup de Atlanta, en vivo.

Todo lo que acabas de leer vive dentro de nuestro Desktop OS — un espacio de trabajo interactivo que se abre directamente en tu navegador. Su mesa de Atlanta mapea la escena startup de la ciudad: quién está construyendo, qué está pasando y dónde. Ábrelo y échale un vistazo.

Abre el Desktop OS →