Remix v3 dropped React. An agent still shipped the app in a day. The team that r
By Mohamadyaraghi · November 13, 2026 · Curated by George's Blog
Remix v3 dropped React.
An agent still shipped the app in a day.
The team that ran the experiment still recommended TanStack Start.
That's the part the "agent-native framework" posts will skip.
This week, a production React shop rebuilt the same real-estate site twice: once in Remix v3, once in TanStack Start. Same backend. Same Tailwind. No UI kit. Claude Code on both, with current docs and the skill file each framework ships.
Remix 3 wasn't even in the model's training data. Four features landed in less than a day. Review found one real mistake.
So yes, agents can learn a new runtime. That's no longer the interesting question.
The interesting question is what they quietly drop.
The Remix agent avoided JavaScript even when users lost a feature. There's no Vite, so Tailwind had to be wired by hand. There's no shadcn. Every React library was off the table. Performance was a wash: first paint was a tie, and the flashy 2.5× "win" was a Lighthouse simulated-throttling artifact that disappeared under real DevTools throttling.
Their conclusion was boring, and therefore useful. They still recommend TanStack Start for most work, because the team already knows React. Remix v3 is for small, standalone apps with almost no ecosystem to miss.
"Built for agents" is a marketing line. Agents will write CRUD on anything if you feed them docs. They will not invent your monorepo, your design system, or the people who already know the mental model.
Don't leave React because a framework is closer to the web. Leave when the constraint is real.
If your team already thinks in React, the agent-native rewrite is the most expensive way to feel modern.