xAI’s Grok Bot is in early beta, and the pitch is blunt: a team of always-on AI teammates that each get their own computer in the cloud, with a desktop, a filesystem, a terminal, and apps. They use the machine like you do, and they never log off.
An agent with a desk
The mental model xAI pushes in its own docs is worth quoting almost verbatim: Grok Bot is an agent with a computer. It has a desktop, a filesystem, a terminal, and apps. That framing sounds obvious until you compare it with every chatbot you’ve used, which had none of those things and borrowed yours only long enough to answer a question.
You define each bot with a name, a title, and a description. One example from xAI’s guide: a bot called Tech Demos whose whole job is to scan your bookmarks each weekday, pick one technology worth demoing, draft a prompt in your style, wait for approve or deny, then kick off a Cursor cloud agent. That’s not a prompt. That’s a job description with a handoff.
What’s in the beta
Grok Bot runs on an xAI-managed cloud VM, uses xAI models, and reaches you through a desktop app with an iPhone companion. It’s early. xAI says so, and the rough edges show: onboarding is sparse, and the fleet concept works better in the docs than in the product some days. But the architecture is the interesting part. The bot keeps state on its own machine, so your automation survives your laptop being closed, your session expiring, and your attention going somewhere else.
Why the fleet idea matters
One agent is a toy. A team of narrow agents with clean handoffs is an org chart. The Tech Demos example ends with the bot kicking a Cursor agent, which is the tell: xAI expects bots to hand work to other bots and to other tools. If that pattern holds, the unit of work stops being a chat and becomes a pipeline you own.
My take
Grok Bot is rougher than OpenAI’s dots and more radical than it looks. Anyone can ship an agent in a chat window. Giving every agent its own computer and expecting them to coordinate is a different bet about where this goes. I’d rather watch a beta try that than watch a polished product babysit a text box.
If you get access, give one bot one boring recurring job and let it run for two weeks. The beta’s rough edges are real, but so is the direction.