xAI has opened a new front in the workplace-agent race with Grok Bot, a beta product that turns Grok into a roster of persistent AI teammates instead of a single chat window.
The pitch is not subtle. A Bot has its own cloud computer, can stay signed in to tools, can work across apps and websites, can coordinate with other Bots, and can keep going after the user closes a laptop. xAI says the beta is available for SuperGrok Heavy, Cursor Ultra, and Cursor Teams Premium subscribers on desktop and iOS, with enterprise access moving through a waitlist.
That puts Grok Bot in direct competition with the workplace-agent products now forming around ChatGPT, Claude, Copilot, Cursor, and other AI work surfaces. The important shift is that xAI is not only selling a smarter answer model. It is selling delegation.
Official Grok Bot video
xAI's launch page includes an official Grok Bot product video showing the teammate-style interface and cloud-computer workflow.
What Grok Bot does
Grok Bot is built around a simple mental model: create a named AI teammate, give it a job, and message it like a person on your team.
xAI's docs define a Bot as a single persistent named agent. Each Bot can keep context, files, browser sessions, preferences, routines, and working history across tasks. The user gives it a role, task boundaries, relevant context, and access to the tools it needs.
The major difference from a normal chatbot is the computer.
Grok Bot runs on a persistent cloud VM with a browser, filesystem, and terminal. It can use connectors and MCP where available, but xAI is also emphasizing computer use for apps and websites that do not expose a clean API. In other words, the Bot is meant to finish work inside the same practical mess of browser tabs, inboxes, documents, CRMs, dashboards, and internal tools that humans use.
xAI gives examples from its internal prototype: sales outbound, marketing campaigns, office operations, bug fixes, CRM updates, invoice processing, and engineering handoffs. The Verge similarly describes it as an agent service designed to complete multi-step workplace tasks from a shared cloud environment, returning when user approval is needed.
That is the key product promise: not "draft this for me," but "go do the work and come back when judgment or permission is required."
How it differs from a workflow builder
The oldest version of workplace automation asked users to map a process in advance. Pick triggers, wire actions, configure APIs, test edge cases, and hope the workflow did not break when an app changed.
Grok Bot tries to invert that setup.
xAI says users can simply message a Bot with a task. The Bot can then work across tools, ask for access, keep the conversation updated, and learn repeatable routines from demonstration. If a user walks a Bot through a workflow once, the product can preserve that path as a routine and run it again later on demand or on a schedule.
That makes Grok Bot closer to a junior operations teammate than a Zapier-style automation canvas. The user still needs to define the result and approval boundaries, but the system is supposed to absorb more of the procedural glue.
The most ambitious part is multi-Bot coordination. xAI says several Bots can share one user-scoped computer, work in parallel, message each other, share context in threads or group chats, and hand off ownership. That opens the door to workflows where one Bot researches, another drafts, another checks, and another files the result.
It also creates a governance problem: if several agents share one computer, they also share files, browser sessions, and app logins unless the account owner structures the workspace carefully.
The Grok 4.6 benchmark context
Grok Bot arrived one day before xAI announced Grok 4.6, and the timing matters.
xAI says Grok 4.6 focuses on long-running agents, ambitious interactive work, coding, research, and multi-step work artifacts. The company says it matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index and published benchmark comparisons across AA Intelligence, GDPVal-AA, DeepSWE, CursorBench, FrontierCode, and APEX-Agents.
Artificial Analysis gives the more useful independent framing. Its Aug. 12 analysis says Grok 4.6 scores 61 on the Artificial Analysis Intelligence Index, in line with GPT-5.6 Sol, behind Claude Opus 5 and Claude Fable 5, and ahead of Kimi K3. The same analysis highlights Grok 4.6's agentic profile: a GDPval-AA v2 Elo of 1753, 50.7% on tau3-Banking, and 88.4% on Terminal-Bench v2.1.
The cost angle is also central. xAI lists Grok 4.6 pricing from $2 per million input tokens and $6 per million output tokens, with a faster variant at twice that price. Artificial Analysis says that keeps Grok 4.6 on the cost-performance frontier for agentic evaluations and puts it below the headline token prices of Claude Opus 5 and GPT-5.6 Sol.
Those benchmarks are not a direct Grok Bot review. They are model-level evidence for the engine xAI wants under long-running agents. The product still has to prove itself in messy real accounts, with permissions, delays, logins, rate limits, wrong assumptions, and human review.
Why this matters
Grok Bot is part of a broader change in AI products: the unit of work is moving from the answer to the assignment.
A chatbot response is temporary. A Bot is persistent. A response lives in one thread. A Bot can own a recurring function. A response ends when the model stops generating. A Bot can continue in the cloud, use tools, write files, and return later.
That shift is why every major lab is racing toward agent infrastructure. For business users, the valuable product is not a slightly better paragraph. It is a system that can handle the weekly account review, reconcile the spreadsheet, prepare the sales list, triage the inbox, reproduce the bug, or draft the launch plan.
If Grok Bot works well, xAI gets something more durable than a model benchmark. It gets a workplace surface where users create roles, routines, approvals, and memory around Grok. That is much stickier than a model picker.
The risk layer
The same features that make Grok Bot useful make it risky.
A persistent cloud computer with browser sessions, files, and app logins is a powerful delegation environment. It is also a place where credential handling, account separation, data retention, and approval design become product-critical.
xAI's docs warn that all Bots on an account share the same persistent cloud computer. Files, browser sessions, and app logins are shared across Bots. The computer is isolated to the user account, not to a single Bot, so a login or file placed there may be available to every Bot that user creates.
That is convenient for handoffs. It is also a reason not to treat every Bot as a separate security boundary.
xAI's approvals docs say users should keep sensitive inputs and consequential actions under control with approvals and secure handoffs. The examples include asking before changing a campaign, deleting or overwriting data, or changing permissions. That is the right framing, but the product will be judged on how hard it is to accidentally over-delegate.
The enterprise questions are obvious:
- Can admins see what Bots did and why?
- Can companies restrict which sites and files a Bot can access?
- How are credentials scoped and revoked?
- What happens when a Bot's memory becomes stale or wrong?
- Can a Bot be prevented from acting on prompt-injected web content?
- How easy is it to separate read-only research from irreversible action?
These are not edge cases. They are the center of the product.
Our take
Grok Bot is one of xAI's clearest attempts to move beyond the chatbot race.
Build Mode made Grok a place to create apps. Grok 4.6 makes the model story more credible for long-running agent work. Grok Bot turns those pieces into a workplace product: persistent named teammates, a cloud computer, multi-Bot collaboration, routines, approvals, and mobile access.
The upside is obvious. A business user could hand off real work without learning a workflow builder or opening a terminal. A founder could create a small staff of AI helpers. A developer could use Bots to reproduce bugs, gather logs, and prepare tickets while staying focused on the main build.
The hard part is trust.
For now, Grok Bot should be treated as an early beta for supervised delegation, not an autonomous employee. The best use cases are high-friction, low-risk tasks where a Bot can gather, draft, reconcile, prepare, and ask before acting. The riskiest use cases are anything involving money movement, permission changes, customer contact, legal commitments, or sensitive internal data without clear approval gates.
If xAI gets the trust layer right, Grok Bot could become a serious workplace-agent surface. If not, it will be another impressive demo that runs into the same wall every agent product eventually meets: the last 10% of work is where the actual responsibility lives.