A computer of its own
It runs in the cloud with a browser and apps: it keeps working with your phone off and the tab closed.
Personal agents · updated 22 October 2026
Agents that never sleep.
First open-source OpenClaw and Hermes, then OpenAI, SpaceXAI, Meta and now Anthropic with Claude Duet: all bet on the same idea, an assistant with a computer of its own that works for you around the clock and only calls you when a decision is needed. Here is what they really do, what they cost and how to use them without handing over the keys.
What sets them apart
Not a chatbot. A collaborator.
A chatbot answers when you write. An agent carries out multi-step tasks. An always-on agent goes further: it has a permanent workspace and keeps going while you are away.
Memory
It remembers preferences, people, deadlines and the thread of past conversations.
Access to your apps
Email, calendar, documents, websites: via connectors or by using the browser as you would.
Initiative
It doesn't just wait for commands: it flags, chases, proposes. It comes back to you for approvals.
The agents
Six agents, who does what.
Grok Bot
SpaceXAI (xAI)
My favourite for work that carries on over time. It is not a single assistant: you create several Bots, each with a role, a memory and its own routines, and they work on a cloud computer even with your laptop off. A main Bot can coordinate the others, so you only talk to one.
- Launch
- 11 August 2026 (beta), on web and in the app since 15 October
- Model
- Grok models, chosen by the platform
- Price
- Included in the X Subscription Premium ($30), Super ($100) and Ultra ($200) a month, excluding tax; Plus does not include it. Existing SuperGrok and Cursor access still applies
- In Italy
- Available on the web and in the apps: check limits from your account
How it works
- Several Bots, several roles: for example one gathers news in your field, one drafts content, one follows the delivery calendar. They work in parallel and pass context to each other.
- It learns by watching you once: it saves the procedure as a skill, and a routine decides when to repeat it (say, every Monday).
- It builds dashboards, reports and memos from the files you allow: the dashboard shows a change, the report explains it, the memo proposes what to decide.
- It messages you on its own when it finds something relevant, and with Grok Voice 3.0 you can talk to your Bots while out walking.
- Marketplace of connectors, plugins and MCP servers; it uses browser, files and terminal like a person.
Watch out for
- Bots on the same account share one computer: files, cookies and logins are common. Give each one only the permissions it needs.
- It signs into your accounts: a phishing email or malicious page can trick it. Keep approvals before sending, publishing or paying.
Sources: xAI — Grok Bot overview · xAI — The X Subscription pricing · TestingCatalog — Grok Bots on web and mobile
Official site ↗Claude Duet
Anthropic
Anthropic's proactive agent, in an app separate from Claude: every morning it sends you a brief of what needs attention across your connected apps, then checks again every few hours and pings you only when needed. When you hand it a task it works on its own cloud computer.
- Launch
- 21 October 2026 (preview)
- Model
- Claude models
- Price
- Waitlist for Pro and Max subscribers; uses your Claude plan limits
- In Italy
- For now in the US, on a waitlist
How it works
- It inherits the plugins and connectors you already use in Claude.
- Before sending a message or paying it shows you the plan and waits for your ok.
- Some actions stay off-limits even if you say yes, like trading or moving money; it never sees your passwords directly.
Watch out for
- It is a preview: Anthropic itself says it can misjudge what is urgent.
- Background checks use up your plan quota.
Sources: Anthropic — Claude Duet · Reuters — Anthropic unveils Claude Duet · MacRumors — Claude Duet
Official site ↗
ChatGPT dots
OpenAI
An agent with its own cloud computer and browser: you give it a goal, connect the apps it needs and decide what it may do alone. It keeps working after you close ChatGPT and brings back results for approval.
- Launch
- 29 September 2026 (DevDay)
- Model
- GPT-6 Astra
- Price
- First dot included in ChatGPT Pro and Business Premium
- In Italy
- Not on Pro (EU, Switzerland and UK excluded); for Business and Enterprise check with OpenAI
How it works
- Used from ChatGPT on web, desktop and phone, by voice too, or from Slack and Microsoft Teams.
- Connects to over 4,000 apps through connectors.
- When idle it does "proactive research", looking for ways to help.
- A read-only mode stops it from using the browser or computer while you are away.
Watch out for
- Expensive: requires Pro or Business Premium.
- Proactivity needs tuning: you decide which alerts and actions are allowed.
Sources: TechCrunch — OpenAI launches Dots · SiliconANGLE — always-on agents with their own cloud computers · Al Jazeera — OpenAI launches dots
Official site ↗
Meta Muse
Meta (Superintelligence Labs)
Meta's personal agent works in a virtual machine with a browser you can watch live: it shops, books and runs long tasks in the background, and tells you when it is done.
- Launch
- 8 September 2026
- Model
- Muse Spark models
- Price
- Free with limits, then $20 or $100 a month
- In Italy
- Not yet: US only, no date for Europe
How it works
- Name, avatar and communication style are customisable.
- A permission system (Sentinel) separates read and write access to services.
- Announced in late September: a real-time video avatar, Mac app control, a dedicated email address, use with Meta glasses.
- Shopping partners such as Stripe, Shopify, Walmart and Instacart.
Watch out for
- Conversations train the models by default (you can opt out).
- Amazon blocked it in September because it did not identify itself as an agent while browsing.
Sources: Axios — Meta debuts Muse · TechCrunch — Everything new coming to Muse · Wikipedia — Muse (AI agent)
Official site ↗Hermes Agent
Nous Research (open source)
The agent that "grows with you": it remembers your projects and, when it solves a problem, turns the solution into a skill it reuses next time. It runs continuously on your computer or a server and you can message it from chat apps.
- Launch
- Open-source project, continuously updated
- Model
- The model you choose (via API or local)
- Price
- Free: you only pay for the model you use and, if you like, a server
- In Italy
- Yes: you install it yourself, wherever you like
How it works
- Persistent memory with a size limit, to keep it tidy.
- Skills created and refined from experience, loaded only when relevant.
- You choose the model, even a local one for maximum privacy.
Watch out for
- You have to install and configure it: some technical familiarity helps.
- Security is on you: you decide what it may touch on the computer or server.
Sources: Nous Research — Hermes Agent · GitHub — NousResearch/hermes-agent
Official site ↗OpenClaw
Progetto open source
The open-source personal agent that made 24/7 agents popular: it runs on your computer, you message it from WhatsApp, Telegram or Discord and it turns a message into real actions, with memory and a skills system to extend it.
- Launch
- Went viral in early 2026, now one of the most followed projects on GitHub
- Model
- The model you choose
- Price
- Free (MIT licence): you pay for the model and any server
- In Italy
- Yes: you install it yourself
How it works
- It connects to the chats you already use.
- Persistent memory and community skills.
- Works with cloud or local models.
Watch out for
- It has broad access to your computer: check third-party skills before installing them.
- Without the right safeguards it is a target for attackers: keep it updated and isolated.
Sources: OpenClaw · Contabo — open-source agent frameworks 2026
Official site ↗| Grok Bot | Claude Duet | ChatGPT dots | Meta Muse | Hermes Agent | OpenClaw | |
|---|---|---|---|---|---|---|
| Where it runs | ✓ (shared by your Bots) | ✓ | ✓ | ✓ (virtual machine) | On your computer or server | On your computer or server |
| Where you talk to it | Web, apps, voice | Duet app, phone and web | ChatGPT, Slack, Teams | Meta AI app and web | Terminal and chat apps | WhatsApp, Telegram, Discord |
| Strength | Coordinated Bots, skills and routines | Morning brief, Claude connectors | 4,000+ connected apps | Shopping and daily life | Builds skills from experience | Big community, many skills |
| Control | Approvals on actions | Plan to approve, forbidden actions | Read-only mode, approvals | Read/write permissions | You configure it | You configure it |
| Entry price | X Subscription Premium | Claude Pro (waitlist) | ChatGPT Pro | Free (US) | Free + model cost | Free + model cost |
Data checked on 22 October 2026. These products change weekly: check availability and prices on the official site before subscribing.
For work
Where they really help.
For a professional or small business the value is not "doing everything alone" but clearing the busywork: sorting, reminding, preparing. The decision stays yours.
InboxSorts, summarises and drafts replies. You review and send.
CalendarProposes times, moves meetings, reminds you of deadlines (tax filings, renewals).
Recurring researchEvery Monday: regulatory updates, grants, supplier prices, industry news.
Repetitive portal choresDownloading documents, filling forms as drafts, checking the status of an application.
Before handing over the keys
Six rules for an agent that won't betray you.
- 01Start read-only
For the first weeks the agent reads and proposes; you write, send and pay.
- 02Approval for anything irreversible
Messages to clients, payments, deletions, signatures: always with your explicit OK.
- 03Dedicated accounts and permissions
Where possible, a separate email and calendar, and only the connectors you really need.
- 04Distrust instructions inside content
An email or web page can hide orders for the agent (prompt injection). A good agent ignores them; you check the logs.
- 05No sensitive client data
For a professional firm: confidentiality and GDPR apply to the agent too. Use business plans with data-processing agreements.
- 06Review every week
What it did, what it spent, where it went wrong. Then widen or narrow its permissions.
My opinion on where this is heading is in "Always-on agents". If you want to see how to set one up safely, I can show you in a lesson.
