All insights

AI agents

Grok Bot: how the new AI agent works inside your apps

The updates shown on October 20 put Grok Bot in the spotlight: an agent that uses a computer and applications to carry out work. Here is how it operates, how to start and where it may suit you better than Claude, Perplexity Computer or OpenClaw.

Grok Bot: how the new AI agent works inside your apps

Grok Bot is the AI agent that can use a computer and applications to carry out work. The updates shown by Matt Palmer on October 20 put it in the spotlight: here is how it works and when it may suit users better than other agents. [1][18]

What the new Grok agent is

Grok Bot offers assistants with a name, a role and context that persists over time. Users can create Bots for research, email, analysis or other responsibilities, then communicate through messages or voice. The product was introduced in beta on August 11; the recent news concerns its evolving capabilities, presented in announcements on October 16 and 20. [1][2][3][18]

How it works: from goals to actions in apps

It works on a cloud computer, with a browser, files and terminal. It uses authorised connectors for supported services and can navigate interfaces for other tasks. Cloud work continues with your laptop closed; your local computer is separate and requires specific permissions. [4]

You specify the outcome, sources and what it may change. The conversation shows operations, created files, questions and approval requests. You can redirect work in progress; an email or message draft can be reviewed before sending. A clear finish line matters: for example, an updated spreadsheet and a set of drafts ready to check. [5]

Each Bot maintains its role and can remember preferences and information from earlier work. Several Bots can coordinate, exchange messages and use shared files: one gathers material, another prepares it and a third reviews it. Memory should still be checked against current sources, particularly when prices, deadlines or customer details change. [6]

A practical e-commerce example

Imagine this request: “Collect today’s support requests, compare them with order status in the authorised services and prepare an urgency summary with draft replies. Stop before sending anything or issuing refunds.” This is an illustrative assignment, not the result of our own test. It defines the outcome, systems to consult and autonomy boundary. The deliverable to review would be the linked summary and drafts, rather than instructions explaining how to do the work.

Getting started and teaching a recurring job

Install the official app, sign in with the account required by the service and create a Bot with a clear responsibility. Add files or connect the necessary tools, then assign a first task. The guide recommends defining the outcome, sources, constraints, deliverable format and review point. A login, two-factor check or CAPTCHA may require you to take temporary control. [7]

Once the method works, save it as a skill, a reusable set of instructions. A routine determines when it runs, on a schedule or after a supported event. Where available, Teach a task lets you demonstrate a browser workflow and obtain a skill to review. In practice, check one report first, then give the same Bot responsibility for preparing that report regularly. [8]

What the latest updates change

The October 20 video brings together improvements to chat cards, speed, draft editing, email aliases, X messages, connectors, voice and notifications. [1] The October 14 changelog also documents attachments up to 250 MB and videos up to 1 GB, more targeted notifications and forms with known fields already filled. These updates arrived over time; they were not all introduced on the day of the video. [9]

Where it may be better than other agents

For X-based work, Grok Bot combines public search without an account with access to messages through the required connection. It may simplify the workflow, without establishing greater accuracy. [4]

Compared with Claude and the capabilities previously called Cowork, the location of actions matters. Claude also continues cloud work with the laptop closed. Its guide, however, requires the desktop app to remain open for driving a browser, using the computer and accessing local files. Grok Bot’s cloud browser can therefore suit tasks that need to keep browsing while your computer is off. The advantage concerns that type of work. [10]

Compared with Perplexity Computer, parallelism and continuous work are not exclusive to Grok: Perplexity already offers subagents, connectors and monitoring. Grok Bot may suit users who prefer organising work around Bots with persistent roles and the X ecosystem. Perplexity also offers fully local execution for some work on supported Macs, with an optional authorised move to the cloud; this may appeal to users prioritising that control over their data. [11][12]

Compared with a self-managed OpenClaw installation, Grok Bot can reduce configuration work through its hosted environment. OpenClaw requires installation, AI access configuration and a running Gateway, while allowing users to manage their infrastructure directly. This is the trade-off between a ready service and a system you administer yourself; it does not establish that either handles every assignment better. [13]

These comparisons describe advantages tied to context. The sources reviewed do not establish Grok Bot as an overall winner on quality, speed and cost through an independent test. A useful selection process is to run the same assignment on suitable products and compare completeness, errors, review time and actual usage.

Access, limits and what to check

Bots on the same account share files and sessions: different names do not isolate access. Auto Review helps assess actions, but the documentation also recommends limited permissions and explicit boundaries. Sending, publishing, purchasing and deletion need appropriate approval controls. These operational details matter because an agent can change data in connected services as well as produce text. [14]

Under the new The X Subscription, Grok Bot is included in Premium, Super and Ultra, drawing on a weekly allowance shared with Grok and Cursor; previous subscriptions follow their applicable terms. [15] Premium costs $30 a month when billed monthly, excluding tax. [16] For accounts based in Italy and elsewhere in the EU, @bot mentions on X are not yet available: this restricts that channel for starting work. [17]

Primary sources

Checked on October 22, 2026. Relevant source sections were read; for the video, the original post and its chapter list were checked, rather than a complete transcript. [1] Matt Palmer, October 20 update: https://x.com/mattyp/status/2112550158885917122 — [2] Initial announcement: https://x.ai/news/introducing-grok-bot — [3] Official October 16 thread: https://x.com/bot/status/2111184270819803351

[4] Computer and apps: https://docs.x.ai/grok-bot/computer-and-apps — [5] Conversations and collaboration: https://docs.x.ai/grok-bot/chat-and-collaboration — [6] Roles and memory: https://docs.x.ai/grok-bot/bots — [7] Getting started: https://docs.x.ai/grok-bot/get-started — [8] Skills and routines: https://docs.x.ai/grok-bot/skills-routines-and-automations

[9] Official changelog: https://x.ai/changelog/bot — [10] Claude/Cowork: https://support.claude.com/en/articles/15520349-use-claude-cowork-on-web-desktop-and-mobile — [11] Perplexity Computer: https://www.perplexity.ai/en-GB/products/computer — [12] Local Perplexity on Mac: https://www.perplexity.ai/en-GB/hub/blog/perplexity-computer-now-runs-fully-local-on-mac

[13] OpenClaw setup: https://docs.openclaw.ai/start/getting-started — [14] Approvals and access: https://docs.x.ai/grok-bot/approvals-security-and-privacy — [15] Grok Bot plans: https://docs.x.ai/x-subscription/grok-bot — [16] Pricing: https://docs.x.ai/x-subscription/pricing — [17] Mention availability: https://docs.x.ai/grok-bot/tag-on-x — [18] Product overview: https://docs.x.ai/grok-bot/overview — Cover: official Grok Bot documentation, docs.x.ai.

Grok Bot: aggiornamento del 20 ottobre, guide ufficiali e documentazione dei concorrentiOpen source