Grok Bot: What xAI Documents, and What Is Already Wrong Online

xAI and Cursor released Grok Bot in beta on 11 August 2026 — a set of persistent agents that sign into your applications and operate them the way a person would. The two companies are, per 9to5Mac, in the process of becoming a single company.

The coverage since has been fast and, in at least one important respect, wrong. This page sticks to what xAI’s own announcement and documentation say, and flags where popular descriptions have drifted.

We have not used Grok Bot. It is gated behind paid tiers we do not subscribe to, and this is a documentation review, not a hands-on test. Everything below is traceable to a source listed at the foot of the page.

What xAI says it is

From the announcement:

Grok Bot is your team of always-on agents. They have their own computer, work inside tools and apps like you do, and keep working 24/7.

Four capabilities are described:

• They operate real applications. Bots “can sign in and work across apps, tools, and websites, including platforms with no clean API or MCP.” That last clause is the interesting one — it is a deliberate bet on driving user interfaces rather than integrations.
• They coordinate with each other. You can “place Bots in a group chat where they can coordinate on their own. They pass work, assign ownership, and only pull you in for judgment calls.”
• They learn a workflow by watching. A Bot can “follow along the next time you do a job. It watches the steps and remembers how you like the work done.”
• They stop for approval. Bots “finish jobs end to end, and only come back when something needs your approval.”

It is in early beta, available to SuperGrok Heavy, Cursor Ultra and Cursor Teams Premium subscribers on desktop and iOS, with Windows and Linux builds and Android listed as coming soon.

The detail most coverage has backwards

The claim circulating in most write-ups is that each Bot gets its own cloud computer.

The documentation says the opposite, in as close to a warning as product docs get:

The computer is isolated to your account, not to an individual Bot. Treat a login or file placed on the computer as available to all of your Bots.

One computer per account, shared by every Bot on it — along with files, browser sessions and app logins. xAI is explicit that this is what makes handoffs work: a Bot can pick up where another left off precisely because it inherits the same logged-in browser.

That design makes the handoffs genuinely useful. It also means the security question is not “what can this Bot reach?” but “what can every Bot on this account reach?” If you sign a Bot into an admin panel to do one job, that session is sitting there for the others. A per-Bot sandbox would give you least privilege; a shared account computer gives you convenience, and those are not the same trade.

That is not a criticism of the design — it is the documented behaviour, and xAI states it plainly. It is a criticism of the write-ups that describe per-Bot isolation the product does not claim.

Claims we could not confirm

Two descriptions circulating widely do not appear in xAI’s material:

Circulating claim	What we found
”Each Bot gets its own persistent cloud computer”	Contradicted. The docs say the computer is scoped to the account, and every Bot shares its logins and files.
”Zero-knowledge multi-agent handoffs”	Not xAI’s language, and it conflates two separate things. Bot-to-bot handoff is a real documented feature. Separately, reporting describes a credential handoff in which you type passwords and 2FA codes yourself rather than giving them to a Bot. Neither is “zero-knowledge” in the cryptographic sense of that phrase.
”It hides its reasoning — no thinking text”	We found no such Grok Bot feature. The nearest documented thing is Think Mode on the Grok chatbot, which does the opposite: it shows step-by-step reasoning.

The credential detail is worth separating out, because it is the genuinely reassuring one and it keeps getting garbled: the pattern reported is that a Bot navigates to a login screen and hands control back to you, you authenticate, and the Bot resumes. If that holds, your password is never handed to the agent. We have not verified it ourselves and xAI’s overview page does not describe it, so treat it as reported rather than documented.

What this is actually a bet on

Most agent tooling of the last two years has bet on integrations — APIs, and more recently MCP servers. That bet is sound where an API exists and worthless where one does not, which is most internal enterprise software.

Grok Bot bets the other way: drive the interface a human drives. It inherits the strengths of that approach, notably reach into tools that will never expose an API, and its weaknesses — interfaces change without warning, and a UI-driven agent has no schema to validate against.

Whether that works is an empirical question, and one week of beta is not enough for anyone to answer honestly, including us.

What to watch before committing to it

• The shared-computer boundary. If you would not give one contractor the logins for every system your Bots touch, the account-scoped computer needs thinking about first.
• Approval granularity. “Comes back when something needs your approval” is doing a lot of work in that sentence. Which actions cross that line is the difference between a useful colleague and an expensive incident.
• What happens when a UI changes. A workflow learned by watching is a workflow coupled to a layout.
• Beta churn. Everything here is dated 18 August 2026 and describes software one week old. Check the source pages before acting on any of it.