A personal agent is an AI agent that belongs to one person, keeps memory about them, works across their apps on a computer of its own, and acts without being asked every time. In the six weeks to 30 September 2026, Meta, OpenAI, Manus and SpaceXAI all shipped one, joining Tencent's WorkBuddy and Alibaba's QwenWork and QoderWake.
They agree on more than the launch coverage suggests. Every one of them gives the agent a computer, a memory, a schedule and an approval prompt. Where they differ is the part that matters once there is more than one person involved: where that computer is, and whether anyone besides the owner can hand the agent work.
This is a map as of 30 September 2026, built from each vendor's own pages. muti is not a personal agent, and where it fits is stated at the end.
Seven personal agents placed on two axes: where the agent runs, from the vendor's cloud to your own computer, and who can give it work, from the owner only to colleagues in a group chat.
What are the seven products?
Meta Muse launched on 8 September. Meta calls it "a personal AI agent. It doesn't just answer questions, it actually does the work." Each user gets a Muse Secure VM, which Meta's engineering post describes as "an isolated linux box with a browser". It is available on iOS, Android, muse.ai and WhatsApp.
OpenAI dots launched on 29 September at DevDay as "remarkably capable, always-on agents built to handle everything". A dot works from "their own cloud computer, their own browser, and the apps you've connected". You reach it from ChatGPT, Codex, Slack and Microsoft Teams.
Manus Cue arrived with Manus 2.0 on 28 September. It is a standalone app for personal agents, "each with its own email, phone number, wallet and computer".
Grok Bot, from SpaceXAI (formerly xAI), launched on 11 August with the line "AI teammates you can give real work to". Each account gets one cloud Linux machine, shared by all of its bots.
Tencent WorkBuddy is a desktop agent that runs on your own computer. You drive it remotely from WeChat, WeCom, QQ, Feishu, DingTalk, Slack, Telegram or Discord.
Alibaba's QwenWork (千问办公) merged QoderWork, DingTalk's Wukong and MuleRun into one product with three layers: a desktop agent, a cloud agent, and a collaboration agent inside DingTalk.
Alibaba's QoderWake runs a team of "digital employees" called Wakers as a local service on your machine. Colleagues reach it with an @mention in a DingTalk, Feishu or WeCom group.
How do they compare?
Why does every personal agent get its own computer?
An agent that only calls APIs stops wherever an integration stops. Give it a browser, a file system and a terminal, and it can work anywhere a person can: sites with no API, apps it has to click through, files it has to convert. Muse writes its own tools inside its VM. Grok Bot's documentation says bots work in places with no clean API or MCP integration. A dot's computer can be opened so you can watch it work.
The cost is credentials. If the computer is in the vendor's cloud, your logins have to go there too. The vendors know this, and their answers are the most interesting engineering in the category. Muse's agent never sees a real credential: it holds "surrogate" tokens, which are exchanged for real ones at the network boundary by a service it cannot touch. OpenAI says a dot can use your saved passwords without exposing them to the model. Grok Bot hands passwords, two-factor codes and CAPTCHAs back to the human, who takes control of the machine to enter them.
The Chinese products started from the other end. WorkBuddy and QoderWake run on the computer you already use, where the files and logins already are. That is also why they need you to leave that computer on: WorkBuddy has a setting to stop it sleeping, and QoderWake's docs say plainly that the device must be online when a task needs local files. The two ends are converging. dots, Grok Bot and Manus 2.0 can each reach your own machine if you allow it. WorkBuddy and QoderWake can each run in the cloud.
Who else can use your personal agent?
This is where the seven split cleanly, and the split follows geography more than technology.
The four American-launched products are built for one owner. Muse has no feature for letting another person talk to your agent, and its small-business version is still single-owner. OpenAI's help centre says other members of a channel can see a dot's messages, but only its owner can direct it. Grok Bot lets your bots talk to each other in group chats, not to your colleagues. Cue puts several agents in a group chat, not several people. Because each Cue agent has its own email address and phone number, strangers can contact it, but that makes it reachable, not shared.
The three Chinese products grew up inside workplace messaging, so they had to answer the question on day one. WorkBuddy's WeCom setup has a visibility scope: "which employees, departments or tags can use this bot". In QoderWake, everyone in a group mentions the same bot, and it routes each request to a Waker. A chat must be approved by an admin, or paired with a code that expires in ten minutes, before anyone in it can use the bot. QwenWork gives each digital employee a name, a department, a responsible person and limits on what it may do.
None of them has a clear answer to the follow-up question: when a colleague's request reaches an irreversible step on your agent, who approves it? The approval prompt in every product goes to the owner about the owner's own request. See how the seven ask permission.
How do they ask permission?
The vocabulary has settled almost word for word. Grok Bot offers allow once, always allow and deny. QoderWake has allow, ask and deny. dots lets you write rules that allow, require approval for, or block an action. Muse's grants run from one-time to permanent. The defaults agree too. Muse "stops for anything hard to undo". Grok Bot recommends gates on sending, purchasing, deleting, publishing and production changes. WorkBuddy and QoderWake block high-risk actions until a human confirms.
Two lines from the vendors' own docs are worth keeping. Grok Bot: "An approval controls the proposed action. It does not reverse work already completed." And also Grok Bot: "Do not use separate Bots as a security boundary", because all of an account's bots share one machine.
Which one fits which job?
Choose Muse or dots if you want the agent to run your own life across consumer apps: shopping, travel, inbox, calendar. Muse is further ahead on consumer commerce and on isolation. dots fits better if your work already happens in ChatGPT, Codex and Slack.
Choose Grok Bot or Cue if you want several agents running routines while your laptop is closed. Grok Bot's routines keep running when you are away. Cue gives each agent an identity of its own that the outside world can reach.
Choose WorkBuddy, QwenWork or QoderWake if the work happens in a WeCom, DingTalk or Feishu group and colleagues need to hand the agent tasks. QoderWake is the most explicit about routing and pairing, and QwenWork is the most tied into DingTalk. Chinese readers can see the detailed comparison, also available in English.
Where does muti fit?
muti is not a personal agent and does not compete with one. It starts from an agent you already run, such as Claude Code or Codex on your own machine, and adds a room so the people you trust can send it work. Each person gets their own permission. Four kinds of action always stop for you, whoever asked: pushing to a shared branch, deleting, spending money and touching credentials. The agent never moves, and your files and keys never leave the machine. See what an agent room is and where agents live in 2026.
Frequently asked questions
What is the difference between a personal agent and a chatbot?
A chatbot answers inside one conversation. A personal agent keeps memory across conversations, holds connections to your apps, works on a computer of its own, and acts on a schedule or on events without waiting for a prompt. All seven products here have those four properties. None of the classic chat apps has all of them.
Which personal agents run on my own computer?
Tencent WorkBuddy and Alibaba's QoderWake run locally by default, and QwenWork has a desktop layer. OpenAI dots, Grok Bot and Manus 2.0 run in the cloud but can reach your own machine if you allow it. Meta Muse runs only in its cloud VM, and Meta has announced a Confidential VM whose key only you hold.
Can I share my personal agent with my team?
Not with Muse, dots, Grok Bot or Cue as of September 2026: they are built for one owner. WorkBuddy, QwenWork and QoderWake let colleagues reach an agent through WeCom or DingTalk groups. If the agent you want to share is a coding agent on your own machine, a room with per-person permissions is the other route. See borrowing an agent.
Are personal agents safe to give my email and payment details?
Each vendor asks before sending, buying or deleting, and the better ones enforce that outside the agent itself: Muse's Sentinel, dots' rules, Grok Bot's per-action approvals. What none of them can do is undo an action you approved. Grant narrowly, and give a spending budget rather than a card with no limit.
Sources
- Meta Muse: about.fb.com · safety design: security.muse.ai
- OpenAI dots: openai.com/index/introducing-dots · help.openai.com · TechCrunch
- Manus 2.0 and Cue: manus.im · cue.im
- Grok Bot: x.ai/news/introducing-grok-bot · approvals docs
- WorkBuddy: workbuddy.ai · WeCom guide
- QwenWork: qwenwork.ai · QoderWake: qoder.com/qoderwake · IM docs
If the agent worth sharing is already running on your machine, give it a room.
Publicado a