Skip to content
El Observatorio

Foundations · 1 min de lectura

When Agents Fail for the Same Old Reasons

Paul Graham's "The 18 Mistakes That Kill Startups" is two decades old and still precise: most startup deaths trace back to not making something people want, and nearly every other mistake on the list is a consequence of that one. For founders building AI-native companies today, the list reads less like history and more like a warning that the new tools do not retire the old traps.

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

A Single Founder Is Still a Single Point of Failure

Graham flags the lone founder as the first and most common mistake because no one person catches their own blind spots reliably. In an AI-native company, this problem compounds. The operating loop of task, action, eval, and human review requires someone to design the loop and someone else to challenge it. When the founder is also the only human reviewer, evals become self-referential and judgment calcifies. The remedy is the same as it was before agents existed: find a co-founder whose disagreement you trust.

Marginal Niches Are Worse When Agents Scale Them

Graham warns against choosing a small market to feel safe. AI-native companies face a sharper version of this trap. Agents can ship volume quickly, but volume in a marginal niche just produces more evidence that nobody cares. The loop runs continuously and leaves behind data that sharpens the next pass — but if the niche was wrong, the loop sharpens the wrong thing faster. Picking the right problem still precedes picking the right architecture.

Platform Dependence Hits Differently at the Model Layer

Graham cautions against building on a platform whose owner can remove the floor beneath you. For AI-native founders, the model provider is that platform. The operating system of the company is agents and workflows, and those agents call APIs controlled by someone else. This does not mean avoiding models; it means designing the harness, the evals, and the human judgment layer so that the company survives a provider switch. What cannot be swapped out is the loop itself and the taste required to run it well.

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 →