In August 2026, SpaceXAI (formerly xAI, now SpaceX’s AI division) launched Grok Bot in beta. The pitch is easy to explain: instead of a chat that answers questions, you get named assistants with a job that work on a cloud computer that stays on, with a browser, files, a terminal and your logins.
Two months later came the news that matters most to anyone deploying agents in a company: Grok Bot started sending the heavy work to Anthropic’s Claude Opus 5.5. The owner of one of the biggest model labs decided the product is not the model. Here is what we pieced together about how it works inside.
Where this comes from
First, a word on the name. There are three “Groks” that do work: the chatbot on X, Grok Build (an open-source terminal coding agent) and Grok Bot. This piece is about the last one.
Who does what
The brand belongs to SpaceXAI, but the product is built and run by Cursor, the AI coding company SpaceX bought for $60 billion in stock, a deal that closed on August 14. Installers, docs and billing all live at Cursor. Grok Bot is not sold on its own: it comes inside Cursor plans (from $120 per seat per month on Teams Premium) and SuperGrok Heavy, with a weekly allowance and usage-based billing after that.
What kind of harness it is
The harness is everything around the model: instructions, tools, permissions, memory and checks. Grok Bot’s is persistent, hosted and general-purpose. Persistent because the computer and routines are still there when you close the app. Hosted because it runs in the cloud, not on your machine. General-purpose because it targets office work, not just code.
- Thin clients. The desktop and mobile apps only send messages, monitor and approve.
- The harness itself. It takes the message, queues it per agent, builds the prompt in fixed sections and runs the model-and-tool loop until done.
- The computer. One isolated virtual machine (Firecracker) per user, with a screen for each bot, a browser, files and a terminal.
- Services around it. Connectors whose tokens are kept off the bot’s computer, MCP, scheduled and event-driven automations, network policy.
- The models. No picker for the customer: Cursor chooses. Since October 7, the main choice is Claude Opus 5.5.
Three ideas that apply to any agent
- Tools on demand. Instead of sending the full catalog with every request, the bot gets the basics and fetches the rest when needed. In one public test, the first request dropped from 40 to 11 tool descriptions.
- Memory in a separate call. After each reply, another model call only extracts what is worth keeping, with its source.
- An action reviewer. A second model checks commands, app calls and screen actions before they run, and decides to allow, ask or deny. Passwords, 2FA and payments always go back to the person.
None of these pieces depends on the model. That is why the engine could be swapped without swapping the product.
Where it is still fragile
Isolation is between users, not between bots. All of one person’s bots share the same computer, files and logins. The reviewer skips memory writes and most settings, and network controls exist only on the Enterprise plan.
For an individual, a shared computer is handy. For a company, it is a purchasing bot that can see the finance login.
Everything runs in the United States, with no on-prem option. For companies with sensitive data and GDPR obligations, that matters.
What it changes for anyone deploying agents
The biggest takeaway from Grok Bot is that the harness is the product and the model is a swappable part. Whoever builds agents into an operation needs to be able to swap the engine when a better one shows up, without rebuilding everything else.
- The AI Analyst maps the routines and defines what each agent may touch and when it must ask for approval.
- The Forward Deployed Engineer builds the setup inside the company: connectors, permissions per team, an action reviewer, memory and logging, with data kept where the law requires.
- Training teaches the team to monitor and approve, which is how trust grows.
Purchasing, support, finance and reporting are just examples. If a task repeats, it can be automated.
Even SpaceX swapped the model and kept the harness. What separates a useful agent from a toy is what you build around it.
Enjoyed this article?
Get the next one before it hits the blog.
A short email per issue with what actually matters about AI agents at work. No spam.