xAI launches Grok Bot, an always-on agent team with its own computers

SpaceXAI introduced Grok Bot on Aug. 11 as an early-beta product built around persistent AI agents with their own cloud computers. A Bot can sign into the apps and websites a person already uses, keep working after the laptop is closed, and return when it needs approval. Access is limited for now to SuperGrok Heavy, Cursor Ultra, and Cursor Premium Teams subscribers; enterprise users can join a waitlist. 1

The difference from a normal chat assistant is the handoff. Users message a Bot with a task instead of first assembling an automation. The Bot can work across tools that have no clean API or MCP connection, learn a routine by watching a person do it once, and run that routine later. Multiple Bots can work in parallel, pass tasks through shared threads, and let one coordinator Bot route work to specialists. 1

SpaceXAI's examples include updating CRM records from call transcripts, processing invoices in Gmail, reproducing a software bug, and handing the fix to another Bot. Those examples show the intended workflow, not an independent success rate. The launch page publishes no model benchmark, model card, or public API specification. VentureBeat likewise reports that the company released no benchmark data for agentic tasks. 12

Grok Bot is less a new model endpoint than a new operating layer for computer use. Anthropic and OpenAI already let agents operate software, but Grok Bot packages persistent execution, learned routines, memory, and multi-agent handoffs behind a messaging interface. VentureBeat places it in that growing agent market and describes its differentiator as the workforce-style packaging rather than proof of a superior underlying model. 2

The access model is also part of the launch. The official product page lists Cursor Ultra at $200 per month and Cursor Premium Teams at $120 per seat per month, with the latter adding team billing, usage analytics, plugins, and SAML/OIDC single sign-on. The same page routes downloads and sales through Cursor, while the announcement presents Grok Bot as a SpaceXAI product. 3

Giving an agent its own computer expands the range of software it can touch, but it also expands the blast radius of a bad assumption. A stale memory, wrong click, or overly broad login could change a CRM record, send a message, or move money. The announcement emphasizes approval when needed, yet the public launch material does not explain reliability metrics, model selection, or detailed permission controls. VentureBeat reports that early testers could not choose the underlying model and that automatic routing was not always strong. 2

For teams already paying for an eligible plan, the sensible test is a sandboxed, reversible workflow with narrow permissions and a human approval step. Everyone else should treat Grok Bot as an important product direction, not yet as a generally available model or API: the next evidence to watch is third-party reliability testing and clearer controls for long-running work.