On August 11, 2026, xAI released Grok Bot in early beta — a service featuring a team of always-on AI agents. Each bot runs on its own cloud VM with a browser, file system, and terminal, and is controlled via regular chat messages, like a colleague. Access is open through SuperGrok and Cursor subscriptions with a separate quota; a build for macOS and an app for iOS are available.
What happened
On August 11, 2026, xAI announced the launch of Grok Bot in early beta. The product is structured as a team of always-on AI agents: each bot is allocated its own cloud VM with a browser, file system, and terminal. Bots are controlled via regular messages, like a colleague; clients are available for desktop and iOS, and state is synchronized across devices. In one chat, you can mention multiple bots: they write to each other, pass tasks and context to one another, and share one user cloud machine. The bot also remembers a process demonstrated by the user as a repeatable routine. Access is included in SuperGrok Plus and Heavy, Cursor Pro, Pro+ and Ultra, as well as Cursor Teams Standard and Premium, with the quota for bot work being separate and not consuming existing plan limits. A v0.28.0 build was released for macOS, and an app in early beta for iPhone. The first public experience was described by the author of the Telegram channel "Iichko": according to him, the agents worked out of the box from a computer, phone, and tablet, and several bots tagged in one chat began writing to each other.
Context
Until recently, agentic systems were most often assembled as a constructor: a developer configured a workflow and connected a separate API or MCP connector to each external service, while acting as a router between applications. Grok Bot offers a different approach — computer use, where the agent itself logs into the necessary applications without a clean API, works with their real interfaces, completes the task, and returns to the human only for approvals. Individually, elements of this design were already known: isolated environments for agents, context exchange between multiple models, learning procedures from a demonstration. The novelty here is not scientific, but product-oriented: xAI has assembled these elements into a ready-to-use user experience and distributes it through already paid SuperGrok and Cursor subscriptions, i.e., through a ready-made distribution channel. The announcement and documentation contain no publications, benchmarks, or evaluation methodology — this is a product release of an early beta, not a research result.
Why this matters for the industry
For the industry, the main shift is in which layer of engineering the vendor takes on itself. xAI itself manages cloud VMs, browsers, and bot coordination, removing the work of assembling and maintaining integrations with each service; the team is left with task setting, configuring approvals, and controlling access rights. Several bots sharing one user cloud machine and passing work to each other remove the human's role as a router between services. The distribution channel is also atypical: agentic teams enter the daily processes of companies — from sales and CRM to bug reports and onboarding — through already established SuperGrok and Cursor subscriptions, without a separate procurement cycle. Startups in the field of orchestration and integrations should reassess: part of the "assemble, configure, and connect services" layer is now closed by xAI itself. Other laboratories also feel the pressure: if xAI publishes reliability metrics, the "agentic team on VM with approvals" format may become a de facto reference for evaluating agentic products, and competition will have to be fought with numbers, not the length of the integration list.
Why this matters for users
If you already have a SuperGrok Plus or Heavy subscription or one of the listed Cursor plans, you do not need to pay additionally: it is enough to install the macOS build v0.28.0 or the iOS app. A separate quota for bot work does not burn the limits of the main plan, so experiments do not take away the usual volume of requests. It is worth delegating a real multi-step task: for example, ask the bot to collect data from Salesforce, filter accounts, and by morning prepare drafts of letters for approval. A process shown to the bot once becomes a routine, which it then performs on a schedule. At the start, it is reasonable to immediately configure approvals and access rights to email, CRM, and files, and to give non-critical tasks first — data collection and filtering, letter drafts, bug reports. Connecting the product to the production environment is still too early: this is an early beta, and it is better to test it on tasks where an error is not expensive.
What is still unknown / limitations
Almost everything known about the capabilities of Grok Bot is vendor statements about an early beta, not measured results. The sources do not contain the share of successfully completed tasks, data on latency, SLA, quota cost, and observability tools; there are no benchmarks or evaluation methodology either. The phrases "the bot completes the task" and "the routine is performed on a schedule" can currently be correctly perceived as product promises. The only live independent review is a post by the author of "Iichko", who himself noted that he had just installed the app and could not draw definite conclusions. It is also not disclosed how the bot generalizes a shown process to new situations, checks the correctness of its actions, and rolls back errors.
Sources
Author
Look at AI, editorial team