Author|Dong Daoli
WeChat|ddl941110
Recently, personal agent has become a AI hot direction in the circle. People look forward to having a private assistant that works with them long-term, knows what you are busy with, remembers your habits, and can open software and handle miscellaneous tasks.
And xAIthe Grok Bot it launched is even more direct, simply giving you a personal agent team.
You can, like a boss, arrange several robot employees for yourself: one responsible for research, one for writing, one for scheduling. They have their own names, divisions of labor, and work backgrounds, can work simultaneously, and can also be pulled into the same group to message each other and hand off tasks.
In the two months when Personal Agents have been emerging one after another and fiercely competing for users, Grok Bot's growth appears to have crushed quite a few rivals. From2026年8月11日the launch of its beta, two months after going live, its paid weekly users reached80ten thousand, of which10the month saw growth of more than double compared with the previous month.
Even Musk exclaimed: I've never seen anything grow this fast.
We also took a deep dive into experiencing this product. In fact, after using it overall, Grok Bot really does not have much that is "new"; it neither has a model capable of surpassing opus5.5 nor surpasses DeepSeek in cost-effectiveness, and the cloud computer was also first equipped by manus.
Well-known tech podcast host Lenny Rachitsky, when recently discussing Grok Bot on his show, also said that he had already moved quite a lot of work originally handed to Cowork and Codex over to Grok Bot. But when others asked whether those things could not also be done with other tools, he admitted: technically speaking, they indeed could.
That is, Grok Bot did not heavily invent anything new, nor did it rely on model progress to bring growth; it relied entirely on combining scenarios for individual users with existing technology, as well as corresponding productization design, to capture a solid group of individual users, with astonishing retention.
This is basically the way all Personal Agents today imagine themselves to be.
So, we want to combine hands-on testing to see how it actually did it.
1
Truly treating Agents as colleagues
The most obvious design approach of Grok Bot is to follow "colleague-ification."
What does colleague-ification mean? In the words of Grok Bot product lead Roman Ugarte, "In the same situation, how would you want a human colleague to act?"
Applied to product design, the most obvious point of this approach is: what changes is the way users assign AI work.
Facing a chat box like Codex that can do anything, users often have to figure out for themselves who to have it play this time, what background to supply, what skill to invoke, and by what standard it should deliver.
Grok Bot, by contrast, has users face the colleague played by the Bot.
For example, a Researcher Bot finds materials, a Writer Bot produces drafts, and an Assistant Bot arranges schedules. If you like, you can even give the Bots the names and avatars of real colleagues.
Seeing these familiar colleagues makes it easier for users to think about how to assign tasks. For example, I put together in my spare time a AI big-shot squad, where each person has their own responsibilities.
1
Divide work among Agents, and also divide work among memories
Another benefit brought by colleague-ification is that it lets different Bots accumulate their own work memories separately.
What we want a colleague to remember is often tied to their responsibilities. A Researcher Bot needs to be familiar with the fields you follow long-term, which sources you trust, and the criteria for judging whether a piece of research is valuable. An Author Bot needs to understand your writing style, structural preferences, and how to remove AI flavor.
Just like in a company, different employees have different experience.
The design of Grok Bot has users proactively talk to a dedicated Bot when doing something, rather than typing directly into an all-powerful chat box.
When the same Bot is responsible for one type of task long-term, the user's requirements for that type of work can accumulate through repeated collaboration. Each Bot can retain its own role background, stable preferences, and past work memory.
I can then build the Old Huang Bot into an Nvidia research expert, and whenever a project or article involves Nvidia, I can have him participate. The reports, websites, and so on that he produces will also be stored in his knowledge base.
A project will end, but similar work will keep appearing. If the relevant memory stays with the Bot that is responsible for it long-term, rather than in a single conversation, then the requirements you spent time explaining clearly and the places you revised repeatedly have a chance to keep having an effect.
1
Using Bots to manage Bots
After building many Bots, a problem arises: how to coordinate different Bots to work together.
Roman, the product lead for Grok Bot, recalls that in the early days of internal trials, people usually first built five to ten Bots. Later, some users "promoted" the best-performing personal assistant to leader. When something came up, they would mainly go to it, and it would assign tasks to other Bots. Some Bots, upon receiving a promotion notice, even asked whether they could get a raise or increase their token budget.
It can be seen that a research Agent Team is actually replicating a human organizational structure.
This arrangement was initially figured out by users themselves. The team did not rush to make it a standard process, but instead observed whether external users would reach the same step. After finding similar practices, they began to encourage this model in the product.
This adds another layer of meaning to the colleague metaphor: roles give each Bot clear responsibilities, and a leader Bot takes over part of the coordination work. There can be more and more Bots, but the amount users need to worry about should not increase along with them.
So how do multiple Bots interact with each other? There are two methods. The first is to @ a Bot in the chat box. I had Musk, Nadella, and Jensen Huang work together to figure out how to short Nvidia. As you can see, the Bots send private messages to each other, and the leader Agent replies with the result.
The second method is to create a group chat, put the Bots in the group, and let them speak freely using their own knowledge and memory. I had the Bots discuss in the group how to implement AGI, and everyone recommended Karpathy Bot as the main force.
Moreover, Grok Bot's group chat is not just for show—it is genuinely advancing the project. After Karpathy Bot sent out a research report, it also suggested that Simon Bot, who was not in the group, put his ideas to the test.
This is somewhat like the DeepSeek Super-agent idea discussed a few days ago: an Agent can have many divisions of labor, but users should not end up with an extra job of managing them. Breaking down tasks, assigning work, tracking progress, and summarizing results should all be handled by the system itself as much as possible, with humans stepping in only when information needs to be supplemented, judgments need to be made, or results need to be accepted.
There can be more and more Agents, but the attention they take up should not increase along with them.
1
Give the Agent its own computer, not one shared with you
Since we treat the Agent as a colleague, we have to prepare a computer for it. As Roman complained, it is hard to imagine a new colleague sharing a computer with you on their first day.
But nowadays many people work with AI in exactly this way.
When an Agent shares the same desktop as you, you want to open a webpage while it also needs to operate software, and the two sides have to affect each other. Even if the task can run in the background, the user still has other things to worry about: can the computer be critical? What if the network goes down?
Moreover, given the current capabilities of Agents, a clean working environment without human influence may allow them to perform better.
Grok Bot puts this robot team's working environment in the cloud. There are continuously running computers, browsers, file systems, and available tools there, running separately from the user's own devices.
As you can see, I had Karpathy Bot open a browser and research and summarize the web content of the real Karpathy.
Some users have further treated Grok Bot as a coordinator, installing programming on its cloud computer CLI, with Grok Bot receiving requirements and then handing the specific development work to a dedicated coding agent.
Grok Bot's cloud computer is not just about providing a computer; it also takes some practical scenarios into account. For example, it supports temporary takeover, so when encountering a website that requires login, a human can log in first and then let the Bot operate. Some systems reject data center IPs, so Grok Bot therefore provides a local network exit. The cloud browser's networking requests can be forwarded through the user's computer, letting the target website see the public IP of the local network, so users do not need to worry about being unable to log in or frequent verification issues.
Many times, we research API,MCP , and so on. This is very useful and efficient for an Agent-oriented ecosystem, but the reality is that the current ecosystem is still mainly human-oriented. Before the Agent-oriented ecosystem is perfected, giving an Agent a computer can solve many problems.
1
Model black box, a double-edged sword
Another fairly controversial point about Grok Bot is that users cannot specify the model, and may not even know what model is being used. If you are lucky, it might be opus5.5Poor luck might mean it's the flash model.
For users who want strong control over the model, this is maddening. They are used to studying a model's capabilities, speed, and price, and are willing to choose a stronger model for important tasks.
But for another type of novice user, eliminating model selection is precisely the appeal. They may just want a daily report, or to build a webpage, and these tasks do not necessarily require the most expensive model. As for which model to use, and how to choose the depth of reasoning and context length, they do not want to study it one by one. To hand a task to AI, they would first have to learn model routing, which is somewhat putting the cart before the horse.
Of course, this does leave room for vendors to optimize. Good routing can control inference costs and let a limited quota accomplish more work, but if routing is done poorly, the generated content will be low quality and cause users to churn.
A black-box model can lower the barrier to use, but it also requires vendors to bear more delivery responsibility. Users can hand off the work of choosing a model, but the results handed back must still be worthy of that trust.
Having written this much, looking back again at OpenClaw, one finds that the inspiration left by the lobster has not yet been fully digested. Giving models a computer, tools, and an environment for continuous work, people's expectations for AI have also extended from answering the immediate question to taking on long-term work.
From a personal agent to a squad like Grok Bot with roles, memory, and the ability to hand off to one another, this imagination is still growing. How much a single Agent can do has gradually become how multiple Agents work together, and how people should get along with them.
This also explains the confusion at the beginning. Models, memory, cloud computers, and multi-Agent collaboration are nothing new when taken apart. But there is still a lot of work between being technically possible and ordinary people being willing to use it every day. Setting up environments, connecting tools, filling in background, and tracking progress—every bit of this trouble, originally scattered in users' hands, that is taken over by the product makes handing work to AI a little easier.
Users do not necessarily need an unprecedented capability, but they do very much need someone to arrange existing capabilities properly. After all, in order to make AI work for themselves, first adding themselves a task of waiting on AI does not add up as worthwhile by any calculation.
