xAI Launches Grok Bot, an Autonomous Agent That Logs Into Your Tools and Does the Work

xAI just shipped Grok Bot, and the pitch is pretty direct: this isn't a chatbot you prompt, it's an agent you assign work to. You give it a task the way you'd brief a new hire, it logs into your actual tools (Zendesk, email, LinkedIn, whatever), does the work, and comes back when it needs your sign-off. The product launched alongside a dedicated announcement post and is available now for download on Linux, with macOS, Windows, and iOS access bundled into existing Grok subscription tiers.

The core idea is agentic persistence. Most AI assistants today are stateless in practice, you open a chat, give context, get an answer, and start over next time. Grok Bot is built to avoid that reset. You spawn a Bot, assign it a role (one on outbound, one on systems, one on a specific project), and it keeps working across sessions, keeping context on how you operate and reportedly getting "smarter over time." Multiple Bots can run in parallel, 24/7, even when your laptop is closed, because the work isn't happening on your machine. It's happening on a cloud computer the Bot controls.

That architecture detail is worth sitting with for a second, because it's the actual engineering story here, not just a UX wrapper on an LLM. Every Grok Bot shares one persistent cloud computer per user. That means your Bots share files, browser sessions, and logins, which is how they hand off work to each other without you re-authenticating anything. Isolation is scoped per user, not per Bot. If you've worked with multi-agent orchestration frameworks, this is a notable design choice: instead of sandboxing each agent separately, xAI is treating the Bot fleet as a single persistent workspace with multiple concurrent actors in it.

How the Workflow Actually Runs

In practice, the demo flow breaks down into a few concrete capabilities:

• Sign-in and tool use: You authenticate a Bot once (the example given is logging into Zendesk to work a support queue), and it operates the interface the way a human would, including tools that are "harder to navigate," which reads as a dig at brittle legacy UIs that don't have clean APIs.
• Routine learning: You perform a workflow once while the Bot watches, and it saves that sequence as a reusable routine it can run autonomously afterward. The source material shows this literally labeled as a "Weekly Reporting" routine being watched and learned.
• Multi-bot handoff: Bots assigned to different functions (Sales Outbound, Chief of Staff, Research, Comms, Travel, Talent Scout, Paid Media, Expense Manager, Product Performance, Bug Reproduction, Account Health) can pass work between each other in the same thread. The example given has a Chief of Staff bot routing a pricing reply by pulling context another bot already had from a prior quarter's thread, answering without waiting on the user.
• Scheduled overnight batches: A Sales Outbound example shows a Bot kicking off at 4:39 AM, researching accounts, scoring contacts on hiring signals and docs traffic, and drafting emails and LinkedIn messages in the user's voice. By 3:47 AM the next cycle, it had a batch of 36 drafted, with 4 skipped because they'd been recently touched, and nothing sent without explicit approval.

That last point is the guardrail that makes this sellable to a sales or ops team: drafts are held for review, and the Bot explicitly states "nothing goes out without me." In the example, a user approves 10 of the 36 queued emails, and the rest stay queued rather than firing automatically. For any team nervous about an autonomous agent hitting "send" on a customer-facing email, that review gate is the load-bearing feature, not the research or scoring.

On pricing, Grok Bot isn't a standalone product so much as a feature tier unlock. It's bundled into Premium ($30/month), Super ($100/month, which gets you 3.5x Premium's usage allowance and priority model access at peak times), and Ultra ($200/month, 12x Premium usage plus earliest access to new models and products). It's also included for Teams and Enterprise plans, and xAI says Enterprise access, including seats for people who don't already have one, requires contacting sales directly.

Security is the other piece worth flagging for anyone evaluating this for a real org. Grok Bot runs on the same SSO, auth, and privacy mode infrastructure as the rest of Cursor's stack (xAI's download links route through Cursor's API, suggesting some integration or partnership there). The cloud computer is encrypted in transit and at rest, with a training opt-out available. Sensitive actions can be routed through what xAI calls "Auto Review" before execution, and Enterprise admins get boot-time controls for DLP (data loss prevention), certificates, proxies, and network policy.

Key Takeaways

Grok Bot is xAI's bet that the next useful unit of AI product isn't a smarter chat window, it's a persistent agent with its own credentials, its own machine, and the ability to act inside existing tools rather than just reason about them. The shared-cloud-computer model is the most interesting technical choice here: it trades per-agent isolation for seamless context handoff between bots, which is a real tradeoff worth understanding before you hand one access to your CRM. The approval gating on outbound actions (nothing sent without explicit sign-off) is doing a lot of the trust-building work, and it's the detail that'll determine whether teams feel comfortable letting these things run unsupervised overnight. Whether "AI teammate" ends up being more than marketing language will depend on how well these routines generalize beyond the curated demo workflows xAI is showing off at launch.