My Experience with Cloud-Based Personal Agents

In recent months, the so-called Personal agent concept has begun to appear on a large scale, with a large number of so-called Personal agent products emerging, represented by Muse、Dot, and Manus also recently launched Manus 2.0 to focus on the personal Agent track, a bewildering array of choices, and I myself have used quite a few. This article aims to discuss my own impressions of personal Agent products and my understanding of this type of product.

How to Define a Personal Agent

Actually, although the so-called Personal agent concept has only been widely mentioned recently, it is currently difficult to define exactly what kind of Agent belongs to personal Agent, because on the one hand, this form of product has actually existed since the emergence of Agent last year (for example, those with cloud capabilities and execution capabilities like Manus), and on the other hand, many Agents not aimed at individuals also have such capabilities, such as Grokbot, which can do far more than what a simple Personal agent can do.

However, overall, most of the Personal agents we see now have the following characteristics:

• There is a cloud virtual machine, usually a Linux system, used to perform operations after the user makes a request, and also supports user takeover to operate;

• The concept of AI conversation sessions is highly de-emphasized, and Dot doesn't even have sessions;

• Mostly uses long-term memory and scheduled tasks to execute daily tasks;

• Usually can connect to the user's real personal data, such as text messages, emails, contacts, etc.

Personally, I think the core of this wave of Personal agents is that it will place greater emphasis on user visibility and the autonomy of the Agent itself. Take Muse as an example; it will check your emails and text messages, analyze the content, and send you messages about necessary emails or text messages (such as subscription upcoming charges, account arrears, etc.), asking how you want to handle them. The visible GUI cloud computer allows you to see the Agent's operations, and users can take over and operate at any time. Unlike Manus, which is actually more like a cloud-based 'black box'—this was almost unseen last year and even in the first half of this year, and it is also the biggest killer feature of this type of Personal agent—obtaining users' real personal information and proactively providing reminders and execution plans.

In summary, I think the currently popular ones that truly belong to Personal agents are mainly the following:

It is not hard to see that I did not classify Grokbot and Manus 2.0's Cue into this category, because in terms of product capabilities, Personal agent capability is only one of the Agent capabilities of such products; however, since these products themselves also possess this capability, they will also be included in the subsequent introduction.

It has highlights, but there are still shortcomings

This may be quite controversial. My feeling is that this type of Personal agent is very interesting now, but it has not yet reached what I would truly consider the state of a Personal agent product in my mind. I would like to discuss this from two dimensions: product positioning and market demand.

The product features are still very limited

This is the biggest problem with these Personal agents today: they look more like a patch applied to a general-purpose Agent to let it obtain some user information and provide proactive notification capabilities, plus a cloud-based virtual machine. I also acknowledged above that this feature itself is quite good, but as a Personal agent it is still far from enough.

Laying out the capabilities of these products side by side, you will find their checklists are highly similar: a cloud Linux virtual machine, a few connectors that can connect to email and calendars, a set of scheduled tasks, and a notification channel that will proactively come to you. The differences mainly lie in whose connectors connect to more services and whose entry point is more prominent, rather than in "what it can actually do for me," that is, the ecosystem.

And the first problem is the ecosystem problem, especially for domestic users. Because of differences in user habits at home and abroad, domestic users in particular will feel that the capabilities of this type of product are insufficient. For example, American users mostly use SMS to communicate with others directly, and email is also a very common way for them to send and receive information. But in China, email and SMS are almost only used to receive verification codes, and more message communication behavior takes place in WeChat and office software. Moreover, almost no domestic company's product is willing to open its ecosystem for Agents to connect directly, just as RSS readers have completely failed to take off in China. This is a very real problem, and it causes a large amount of important information in the domestic ecosystem to be inaccessible to AI, which is a very awkward problem.

In mid-October, Doubao Mobile Assistant published a refusal list: more than 150 mainstream apps explicitly stated that they do not allow it to perform screen automation operations within their interfaces. WeChat, Taobao, Meituan, Xiaohongshu, and Kuaishou were all on it; the Tencent ecosystem alone accounted for 66 of them (original link:155 apps refuse to integrate with Doubao! AI taking over agent phones faces challenges__Nanfang+_Nanfang Plus). Another path is called "going through the front door": WeChat partnered with Huawei, Honor, Xiaomi, OPPO, and vivo to launch A2A capabilities, but the open list has only three items—sending messages, making voice calls, and making video calls. It cannot read chat records, cannot access Moments, and cannot manage group chats, and the execution rights and risk control rights always remain with WeChat (original link:From banning Doubao to embracing Honor, why did WeChat suddenly "change face"? - TMTPost official website). For a Personal agent, these three capabilities are roughly equivalent to "being able to press the send button for you," and are still far from "knowing what you have been busy with recently."

Moreover, domestic users have an even more troublesome issue: the data itself is fragmented. Take users in the United States overseas as an example. Their life records almost always end up deposited in one email inbox: bills, itineraries, verification codes, notifications—any service that wants to contact you ultimately has to send an email, so integrating with one Gmail is equivalent to obtaining a running record of your life. That is not the case in China—the electricity bill you pay in WeChat is stored in the power grid company's system; the appointment you book in a mini program is recorded on the hospital's server; ride-hailing, food delivery, and ticket booking each remain in their respective apps. This means that in China no product can let users see all data at once in just one product, and the result is that there is no shortcut of "integrate once and that's enough."

Of course, overseas looks more relaxed, but it is not actually that taken for granted. Muse had been online for only a little over ten days when Amazon blocked its purchasing-agent function on the grounds of "unauthorized AI agents continuously accessing" the site; sites like Target directly intercept automatic clicks at checkout, and six global banks jointly issued a statement warning about the risks of AI-operated actions. Manus simply changed its approach in Cue (although it may also have borrowed from Grokbot): instead of connecting to your account, it gives each Agent its own identity—an independent email, phone number, wallet, and computer—so that it can send messages, make payments, and handle tasks as a "newcomer," thereby bypassing a large number of authorization troubles. But this means that the Agent itself is not Iron Man's Jarvis; it can still only rely on an independent identity to do things, rather than truly doing things for you.

The second problem is that the complexity of using features is still too high. Many features require users to actively explore, discover, and enable them, which actually goes against the characteristics of consumer products. In my experience, the best product so far is Muse. Muse made configuring the modules that needed to be connected almost imperceptible to me. Other products, although they do not actively ask me to configure things, also do not tell me where the configuration is, which easily leads users to treat them as dispensable chat products.

This complexity has a concrete shape. Taking Muse as an example again, connectors are hidden in the Connectors section in settings, and the official suggested order is "first connect calendar with read-only access, then connect email with read-only access, and only consider opening write permissions after you are used to it" (original text:blog.einverne.info). This sentence itself assumes that users are capable of distinguishing between "read-only" and "write." However, this is all a relatively ideal state. In actual use there are still a bunch of pitfalls. For example, two-step verification codes have to be entered manually, and for security reasons the email connector actively filters out verification codes and login links; disconnecting a connector only stops it from fetching new data, and the content it has seen before may still remain in memory—if you want to clean it out completely, you have to specifically make it forget. These are common knowledge to developers, but for ordinary users they are a threshold, and consumer products should least of all have thresholds.

Perhaps AI's capabilities are too strong, which has led everyone over the past year or so to always want to build something omnipotent, gradually making products general-purpose and all-capable. As a result, they have instead lost the exploration of vertical capabilities that deeply address user needs as a product. This is like the "curse of knowledge," making the product managers, designers, and developers of every AI product begin to feel that users only need to open settings and configure things. But we have ignored that many users using Doubao do not even know how to start a new conversation, or even that they can view historical conversations. Products in this form will ultimately only become the self-celebration of a small number of developers.

The bumpy revenue road for AI in the consumer sector

The revenue potential of the consumer segment is currently very concerning. We all know that Doubao has so far enabled paid features, but the actual paying user base is negligible, and it is still at a stage where revenue is far less than cost, let alone recouping the investment. Looking abroad, companies like OpenAI and Anthropic make profits on one hand by selling API quota to B2B customers, and on the other hand lose money on the consumer market by subsidizing subscriptions to burn through usage. How exactly the consumer segment should make money—currently the market has no perfect answer.

But Personal agents also carry some higher cost burdens—the cost of an ordinary chat product is basically equal to one inference; if the user doesn't ask, no money is spent; but the cost of a Personal agent is always-on: a cloud virtual machine running 7×24 hours, a bunch of scheduled tasks, writing and retrieving memories, and continuously polling emails, text messages, and calendars in the background. The user may say only three sentences to it in a day, yet it runs full 24 hours in the background. What's more troublesome is that the tasks it takes on are generally long tasks and multi-turn tool calls; a complete task often involves dozens of model calls, and token consumption is more than ten times that of ordinary conversations. This means that the marginal cost of a Personal agent does not decrease with scale; instead, it increases as the user's level of trust rises: the more the user relies on it, the more it runs, and the more you lose. This runs contrary to the classic internet product logic of 'scale dilutes cost.'

Now for revenue. The 'subscription budget' of consumer users is limited: video memberships at 20 yuan, music at 15 yuan, cloud storage at 10 yuan—the mental account has long been filled up. When a new assistant wants to charge 20 US dollars a month, the user's first reaction is often not 'what can it do for me' but 'this money is enough for me to order takeout several times.' What's even worse is that the value of this type of product is implicit: it helps you avoid an automatic renewal you forgot to cancel, helps you reply to a difficult email—but users' perception of this value is weak, and it is hard to charge for it; yet if it makes a mistake once—deletes an email that should have been sent, pays a bill that should not have been paid—the loss is explicit and attributable, and trust collapses extremely quickly. Value is hard to price, while errors are amplified—this is the most awkward aspect of consumer AI products.

Perhaps the more realistic options are these few:

• Charge by result. For tasks with a clear transaction loop, such as booking airline tickets, booking hotels, placing orders on behalf of users, and handling refunds, charging a commission based on completed transactions is much more natural than time-based subscriptions;

• Charge by usage. Convert inference and virtual machines into credits, deduct as much as is used, and heavy users bear the cost themselves;

• Turn it into an entry point for traffic and transactions. Agents naturally stand at the very front of 'what does the user want to buy'; recommendation, price comparison, and discount distribution are themselves a business;

• Let the B2B side support the consumer side. Make the personal version free as an entry point and for reputation, and make money from team versions and enterprise seats—this has been verified many times in collaboration tools.

But frankly speaking, none of these is a 'perfect answer.' They all require the product to first have sufficiently deep scenarios and high-frequency usage, while today's Personal agents are precisely insufficient in both. In the final analysis, business problems are ultimately pushed back to product problems.

How Personal agents may develop in the future

Earlier I said the ecosystem is the biggest problem, but looked at from another angle, this is precisely also the biggest opportunity for companies in this track—because where the data is, that is the only place a Personal agent can grow.

Just now I said that Personal agents in the United States can take off because infrastructure such as text messages, email, and calendars is open and programmable, whereas in China real personal information is not in these places—it is in WeChat, Douyin, Feishu, Alipay, and in the phone system itself. So China's Personal agent will most likely not be yet another standalone App, but will grow on the party that holds these entry points. Still, whether they will ultimately really capture this market is unknown. But at least we know it must be a product with a simple conversational entry point and a cloud-based form, and it may even be somewhat aggressive in obtaining data and permissions.

Jarvis still has a long way to go.

A personal agent is an agent I'm willing to hand over part of my life to. But right now, it faces the same problem as agents in other professional scenarios today: does it really know what's happening in my life, and do I dare let it make decisions for me.

As for whether this wave of products will ultimately break through, I don't think we'll be able to tell within this year. Gaps in capability will quickly be closed; what will really set them apart is who first pries open a crack in that wall of the ecosystem, and who first makes users feel that "it's okay to hand this over to it." The former is a resource problem, the latter a product problem—and the latter is often harder.