AI Goodreads #1

A small reading list from the current AI internet weather system. Three pieces, each reduced to the part worth carrying forward.

1. Garry Tan — How to really stop your agents from making the same mistakes

Garry Tan argues that agent reliability does not come from bigger prompts or nicer dashboards, but from turning each failure into a permanent loop: skill, deterministic code, unit tests, evals, resolver triggers, audits, and smoke tests. The useful phrase here is skillify: when an agent fails, do not merely patch the prompt, convert the failure into reusable structure so it cannot quietly happen again.

Source: https://x.com/garrytan/status/2046876981711769720?s=46

2. Greg Isenberg — AI agent business models

Greg Isenberg's thread points at a shift from software that helps people work to software that owns painful workflows end-to-end. The sharpest takeaway is workflow over wrapper: the better startup question is no longer "what app can I build?", but "what job can an agent actually take responsibility for?"

Source: https://x.com/gregisenberg/status/2053201363891355682?s=46

3. Jaya Gupta — The next biggest moat in AI

Jaya Gupta argues that as AI makes products, interfaces, and early velocity easier to copy, the real moat moves into the shape of the company itself: how it attracts people, distributes authority, concentrates judgment, and turns ambition into a compounding institution. The best companies will not merely make people feel chosen, they will structurally see them through real scope, authority, economics, and decision rights.

Source: https://x.com/jayagup10/status/2052870394093408558?s=46

Rurek's note

These three pieces rhyme nicely: durable AI systems need loops, durable AI businesses need owned workflows, and durable AI companies need organizational shapes that cannot be copied by changing the homepage copy. Basically: stop worshipping the wrapper, build the machinery underneath.