X.com rewritten from scratch in six months by 5 engineers and AI agents
A team of fewer than 5 engineers used AI agents to rebuild X.com in six months on a single server powered by React Suspense, Relay, Fastify, and Vite.

The X engineering blog published a detailed breakdown of how they migrated the web app to a brand-new stack. Previously, they ran a patchwork of three separate servers: a client-side rweb server (built on an aging stack of Redux, Flow, and React Native for Web) for human users, and two separate SSR servers (rweb SSR and Jetfuel SSR) for search engine crawlers. Any tweak to SEO or meta tags turned into a headache because the servers kept drifting out of sync.
Everything now runs on a single server that tailors responses based on who is making the request:
- Regular users get streaming HTML via React Suspense. Pre-fetched GraphQL payloads stream alongside the markup so the browser does not have to fire extra requests during hydration. On slow 4G, time to first post dropped from 23.9 seconds to 2.7 seconds.
- Search engine crawlers (verified strictly by IP rather than
User-Agent) receive fully rendered static HTML without the client bundle to stay well within Googlebot’s 2 MB limit. - Link preview bots (Telegram, Slack, Discord, etc.) get only the
<head>tags without rendering the page body. The preview card images themselves are now generated on the fly by a dedicated Rust microservice (rewritten from a JS service that used to stall for up to two minutes under load).
What is under the hood of the new stack:
- Build and runtime: Vite (powered by Rolldown), Fastify, and Platformatic Watt running on Node.js v26 (which manages multiple workers in a single process and automatically restarts any worker with a blocked event loop).
- Frontend and data: React, Relay + GraphQL, TypeScript, and Tailwind CSS alongside the new XDS design system built on Base UI. Typechecking runs on
tsrs(the Rust port of the TypeScript compiler). - Navigation: code and data prefetching on hover whenever the cursor rests on a post for more than 100 ms.
How they shipped it so fast:
The initial version was built in six months by a team of fewer than 5 people leaning heavily on AI agents. Engineers designed the architecture and reviewed PRs while agents ported business logic from the legacy servers and wrote tests.
To keep the AI from generating slop, the team enforced strict guardrails in the repository:
- A single root
AGENTS.mdwith all repository conventions (no barrel files, named exports only, strict boundary validation). - Over 25 reusable skills for routine workflows, from working with Relay to running CI.
- Custom
oxlintplugins, including a dedicated “anti-slop” linter that automatically bans common LLM patterns likeunknownparameters, conditional object spreads, and chained type assertions.
Logged-in X.com is now rolling out country by country via lite.x.com, while logged-out traffic and crawlers have already fully migrated to the new architecture.