Personal Agent究竟是不是一门好生意？|甲子光年

Is Personal Agent Actually a Good Business? | 甲子光年

Have you used Personal Agent?

Author | Lizi

“Exactly how many ordinary people need a real‑life assistant?”

In October this year, at a closed‑door meeting on Personal Agents attended by “Jiazi Light‑Year”, someone asked this question.

No one present answered immediately.

Because behind this question lies another simple question:

Isn’t a Personal Agent just an app that helps you book tickets?

Even though this issue may seem amateurish in today’s AI-dominated era, when you try to answer it seriously, it really seems that no one can provide a solution that satisfies everyone.

Everyone has a different understanding of a Personal Agent.

Some say it’s a private assistant that can book flights, write emails, and cancel subscriptions for you; others say it’s a digital twin with memory and personality that can proactively help you; and some say it’s just a more powerful OpenClaw.

Although there is no standard answer, Personal Agent has already become the hottest track worldwide.

Meta’s Muse reached 5 million downloads in three weeks; Instinct, with a 14‑person team and a $10 billion valuation, raised $1 billion, and its beta invitation was being resold for $300 on eBay.

According to the personal AI assistant evaluation site Assistant Benchmark, as of September this year, it has cataloged 122 personal AI assistants, of which 64 are general‑purpose, accounting for more than 50%.

On one side there is confusion, on the other side there is celebration.

This is the most realistic situation of Personal Agent right now.

In today’s tech circles, where everyone talks about Personal Agent, “Jiazi Light-Year” wants to understand: what exactly is a Personal Agent? Why is it so popular now? Who is actually making money in this arena?

In a nutshell: Is a Personal Agent actually a good business?

1. What exactly is a Personal Agent?

The reason there are a thousand Personal Agents in a thousand people’s minds is because Personal Agent has not yet settled on a unified product form.

For example, abroad Muse and domestically Today are both standalone apps; Instinct doesn’t have an app at all—the interaction is via phone calls and text messages; while OpenAI’s Dots is hidden inside ChatGPT.

Dots product screenshot Image source: Jiazi Light‑Year

Although the product form of Personal Agent has not yet converged, the industry has basically reached a consensus on its essence:It is a silicon‑based agent centered on a person’s long‑term state, with a dedicated cloud runtime environment and persistent memory, capable of proactively completing cross‑application closed‑loop tasks for the user within permission boundaries.

Compared with all previous AI products, the Personal Agent has at least three fundamental changes.

The first change is moving from “outputting answers” to “delivering results.”

In the past, traditional AI assistants based on large models strictly followed a passive, responsive loop of “input Prompt, output text result,” and could not perform actions for users such as price comparison or form filling.

The Personal Agent, however, delivers end‑to‑end execution results and consumes more compute power.

According to estimates, a traditional chatbot consumes only about 8K tokens per conversation on average. In contrast, a Personal Agent, which performs multi-step decomposition, cross‑web trial‑and‑error, and verification in the background, consumes over 256K tokens for a single complex task—32 times more than the former.

The second change is that the state center shifts from “task” to “person”.

A simple question is: why haven’t office‑oriented AI agents like Codex and Claude Code, which already possess strong autonomous execution capabilities, directly evolved into today’s Personal Agent?

Former TikTok product manager Mo Weishu pointed out: “The core distinction between a Personal Agent and an office Agent lies in its context—it’s built around the person rather than around the task.”

Strictly speaking, the core users of products like Codex and Claude Code are developers, so their interfaces are designed around projects, code repositories, and workspaces to obtain dedicated context, with the goal of completing various development tasks.

But ordinary people’s daily lives often involve highly fragmented scenarios such as changing flight tickets, booking a restaurant for an anniversary, or canceling a membership they forgot to unsubscribe from. This means the context for a Personal Agent must be built around the user themselves, putting the person at the center.

The third change is a dramatic reduction in the barrier to entry.

In fact, the cross‑application capabilities showcased by today’s major Personal Agents were already taking shape in the early‑year sensations OpenClaw and Hermes. So why did it take another half a year for a true consumer‑level explosion to occur?

“The main reason Muse became popular is that it lowered the user interaction barrier to an extremely low level.”

Not long ago, the founder of K2Lab (Panfeng Intelligence), Wang Ming, who just completed a multi‑million‑yuan angel+ round from Kaitai Capital (the third round within six months), analyzed “Jiazi Light‑Year”.

“We all think Codex is already very useful, but for the general public the barrier is still too high. Muse is like a premium version of Codex that anyone can use; the interface is ultra‑minimalist, interaction is convenient, and the underlying Computer Use and Browser Use are robust, allowing it to work for you online all day long,” Wang Ming said plainly.

Muse product interface Image source: Apple Store

The decisive driver that underpins the above changes and sparks the popularity of Personal Agentsis the concentrated availability of underlying technology this summer.

For example,the success rate of long-chain operations is higher than before.

According to Leah, the founder of Karpo, a Personal Agent startup focused on urban living experiences, the execution success rate for long tasks in previous generations of models was roughly 60%. This summer, with the wave of Fable 5 and GPT‑5.6 Sol, the success rate has climbed to over 80%.

For example, the widespread adoption of cloud‑based virtual machines and independent digital identity infrastructure.

To address the early‑stage Agent issue of occupying the user’s local device screen and being interrupted when the device is shut down or loses connectivity, products like Muse and Dots have opted to provision dedicated cloud‑based virtual machines to ensure continuous task execution.

In addition, many products have started issuing each Agent an independent digital badge, equipping Agents with dedicated email addresses, phone numbers, and even encrypted wallets that support quota‑based authorizations, giving AI a legitimate identity to act independently in the digital society.

Obviously, once the product and technology were fully ready this summer, the rise of Personal Agents became inevitable.

But the question is, why does Personal Agent seem to be booming in the United States while China has not experienced the expected explosion?

2. The U.S.’s genuine essential demand, China’s pseudo demand?

Personal Agent has sparked a phenomenon-level consumer frenzy across the ocean.

According to estimates from app analytics firm Apptopia, Muse recorded about 642,000 daily active users on U.S. mobile devices in its first 12 days, nearly three times the performance of ChatGPT during the same period that year.

Muse download data Image source: Apptopia

In contrast, Instinct, which has no graphical interface and relies entirely on SMS and phone interactions, is even harder to find in Silicon Valley.

A widely circulated story goes like this: a couple from the United States, while watching a match at the US Open in person, sent their seat number to Instinct, asking it to help locate the camera shot that showed them on the venue’s big screen.

Instinct apparently spent 18 hours behind the scenes—sending emails, checking recordings, contacting customer service, and finally tracing the responsible US Open video editor to obtain the video file.

However, in China, Personal Agent has not yet sparked the OpenClaw‑style hype that emerged earlier this year.

Many practitioners believe thatPersonal Agent may be a false demand in China.

This stark contrast is actually not hard to understand.

First, the U.S. internet ecosystem is highly fragmented.

It may be hard for us to imagine that, even today, the U.S. internet still retains a highly decentralized web ecosystem.

An ordinary person in the U.S. often has to hop between dozens of separate websites, apps, and email to handle payments, food orders, tax filing, and other needs, and some local small businesses and community clinics even accept only phone calls and text messages for appointments.

According to a cross‑national administrative burden survey cited by the media, U.S. adults spend an average of 11.8 hours per month handling administrative chores such as medical claim verification, benefit applications, and tax filing. Across the United States, just filling out federal government forms consumes more than 12 billion hours each year.

In such a highly fragmented scenario, a Personal Agent that only needs an email address, knows how to use a browser, and can queue phone calls on behalf of the user can indeed unleash massive information‑integration value instantly.

China, however, is completely different.

In China, the evolution of mobile internet has already penetrated deep into daily life. Paying utilities, ordering groceries, handling government services, and arranging travel and hospitality have long been highly integrated into a few super‑entry points such as WeChat, Alipay, and Meituan.

“The domestic mobile internet has evolved so thoroughly, creating a “frog‑jump” effect that diminishes AI’s presence,” K2Lab founder Wang Ming said bluntly.

He cited the AI milk‑tea battle at the beginning of the year: “After trying it, you’ll find that ordering milk tea with Qianwen AI versus the user opening Taobao Flash Sale and placing an order—both in steps and implementation—do not create a noticeable experience gap.”

Secondly, the degree of ecosystem barrier closure also varies.

Overseas, users’ bills, flight confirmations, and calendar invites ultimately converge in personal email, and the web remains the primary service carrier.

In China, users’ personal context and transaction loops are locked within the apps of major giants, and the data wall remains difficult to break.

Moreover, once someone tries to cross ecosystems and touch another company’s interests, they will inevitably face retaliation.

On October 14, the official Doubao Mobile Assistant disclosed the latest progress of the “Screen Automation Operation Statement Agreement” (SAEP) and the list of refused operations. High‑frequency apps such as WeChat, Meituan, Alipay, and others are included, with more than 150 applications explicitly refusing to integrate with Doubao Mobile Assistant.

During its visits and exchanges, “Jiazi Light‑Year” discovered that many practitioners believe the opportunity for a Chinese Personal Agent may only appear within ecosystems like Alibaba, Tencent, etc., which can close the loop on all life needs.

There is another hidden difference: the labor cost versus users’ willingness to pay for “time‑saving.”

According to Wang Ming’s observation, a key reason Personal Agents thrive overseas is “laziness.”

“The U.S. has dozens of e‑commerce platforms, and labor costs are extremely high. Cross‑platform price comparison is a real need, and people are willing to pay for it. Chinese users, however, are too diligent; whenever they can find a way to save a few bucks, they do it themselves without needing AI, let alone paying for it.”

3. General-Purpose Personal Agent Is a Money‑Burning Competition

Since overseas market demand truly exists, has the universal “do‑everything” Personal Agent already closed the commercial loop?

The answer may be exactly the opposite.

According to observations by “Jiazi Light‑Year”, whether it’s Muse, whose growth rate far outpaces ChatGPT, or the hard‑to‑obtain Instinct, their business models are under pressure.

The biggest problem is that Personal Agents consume too much compute power, which is the root cause of everything.

Currently, virtually all general-purpose Personal Agents claim they can help users fulfill all kinds of requests around the clock. To maintain a 7×24 cloud presence, sustain an independent virtual machine, scan email calendars every few minutes, and perform repeated multi-step reasoning on web pages, their computational consumption inevitably expands exponentially.

According to Barclays analyst Ross Sandler, the operating cost of a free Personal Agent is roughly ten times that of a free ChatGPT service.

And based on a16z’s estimate, a fully utilized Personal Agent could cost as much as $20 per user per day. This means serving 100,000 active users would burn hundreds of millions of dollars in compute power annually.

However,compared with the massive computing costs, consumer subscription fees are merely a drop in the bucket.

a16z conducted a survey of 44 AI-native products: 84.1% offer subscriptions, 63.6% charge based on usage or credits, 13.6% have ads, and 2.3% rely on platform transactions. In other words, more than 80% of AI products primarily use a subscription model.

Consumer AI pricing model statistics Image source: a16z

For example, OpenAI bundles its own Personal Agent product Dots into the $100‑per‑month ChatGPT Pro subscription.

But the question is, are there really that many people willing to pay a hundred dollars a month just to save ten minutes of ticket‑booking time?

According to a16z’s analysis of YipitData’s U.S. consumer electronics receipt sample, as of August this year only 4.5% of U.S. consumers had subscribed to at least one of ChatGPT, Gemini, or Claude, with a median monthly spend of just $25 among paying users, and no clear long‑term growth.

Another data point comes from a July survey by venture firm Menlo Ventures of 5,067 U.S. adults, which found that 64% of Americans have used AI. Among those who have used AI,only 14% are paying more than $100 per month (eligible for Dots).

Since it’s hard for the product to charge subscription fees to end users, Personal Agent chose another highly mobile‑internet‑styled business strategy: free for consumers, charge enterprises.

Typical examples are Muse and Instinct.

Muse’s approach is free basic features, with heavy usage requiring a subscription.

Instinct’s strategy is even more internet‑like: currently free to use, it grows transaction volume by saving users money and handling purchases on their behalf, then takes commissions from hotels, service providers, and other B‑side partners.

Just looking at the front‑end growth and transaction volume, Instinct indeed hands in a seemingly decent report card.

According to Instinct founder Noah Shinn himself disclosed on a podcast, 40% of Instinct users linked a credit card within three weeks, and among users who have authorized at least one piece of sensitive information, the retention rate is as high as 80%.

Even more noteworthy, once these users make a purchase, they spend over $1,300 per month on average through Instinct. Based on this data, with just over 100,000 seed users, Instinct’s annualized transaction volume has already exceeded $1 billion.

But even so, Instinct has never opened registration, and the core reason remains computational cost.

Noah Shinn admitted on the podcast that under the current invitation system, Instinct’s compute costs double every week. Because compute delivery has a lag, the company cannot wait until it’s needed to purchase; it must procure in advance.

This means Instinct is actually carrying a high financial leverage, and a mis‑forecast could result in huge losses.

On the other hand, charging the B‑side may not be an easy path either.

So far, the value of overseas‑wide Personal Agents has been built on aggregating fragmented information, using methods like whole‑web price comparison to facilitate transactions and taking commissions from merchants. But this obviously directlyhits the cheese of e‑commerce and local service providers.

On September 20, Muse was completely banned by Amazon less than two weeks after its launch.

The reason is simple: Amazon’s internal advertising revenue alone reached $68.6 billion last year. If external agents were allowed to let users skip waterfall ads and place orders directly through whole‑web price comparison, the e‑commerce giant’s advertising business model would collapse instantly.

4. Who Is Actually Closing the Business Loop?

Since even star projects have yet to close the business loop, does that mean the hot Personal Agent is just a “seemingly beautiful” business?

Not necessarily.

From the perspective of “Jiazi Light‑Year,” a general‑purpose Personal Agent is unlikely to break even with a modest subscription fee due to the high compute costs, so it will inevitably become a battlefield where giants and well‑funded star startups burn cash to fight for the next AI entry point.

For many startups that want to establish themselves in the Personal Agent arena, “vertical specialization” may be the optimal solution.

According to Wang Ming, founder of K2Lab, if a startup chooses a super‑general entry point, it will inevitably be crushed by the big players.

“From day one we made it clear: we will only build Personal Agents for vertical audiences and focus solely on delivering direct commercial outcomes.”

In fact, there are already many niche Personal Agents emerging in the market.They achieve a commercial loop by focusing on a specific user group and solving a specific need.

For example, K2Lab is a representative example that specifically helps people make money.

K2Lab co‑founder Tang Minglei divides Personal Agents into four quadrants: saving time (efficiency), killing time (entertainment), saving money, and making money.

As a typical case of a “Personal Agent that helps people make money,” “Jiazi Light‑Year” has repeatedly tracked and reported on K2Lab, dissecting its business model.

In short, K2Lab helps content creators on overseas video platforms earn money through product promotion and shares profits with them on a CPS (cost‑per‑sale) basis.

K2Lab’s approach is very straightforward: they target the huge production vacuum in overseas TikTok content e‑commerce.

TikTok boasts over $1 billion in daily active users worldwide, yet the vast number of mid‑tier creators and ordinary users abroad (such as full‑time moms and part‑time office workers) neither know how to select products, write scripts, edit, nor run ads, and they strongly dislike the MCN model’s lengthy, unrewarded training and long‑term contract binding.

To address this, K2Lab has created a Personal Commerce Agent for overseas creators—Moras—that handles the entire chain from hot‑product selection, script creation, multimodal shoppable video generation to publishing and performance review.

Previously, an unemployed mother of two young children who had been laid off by the U.S. Postal Service generated $14,000 in GMV within less than a week of trying Moras, and her monthly sales even surpassed the $100,000 mark.

In terms of business model, K2Lab has achieved a seemingly impossible leap in the past six months.

“Jiazi Light‑Year” learned that during the co‑creation phase with a hundred seed users at the end of April, the team was still testing a pricing threshold of “$20 monthly subscription fee plus commission.”

However, after the product officially launched around June‑July this year, K2Lab eliminated the $20 subscription fee and fully adopted a “Zero‑Three Model” (zero subscription fee, zero token fee, zero MCN binding agreement), implementing a pure CPS‑based payment system!

That means overseas influencers can use the full suite of AI capabilities for free without spending a cent, the account always remains personally owned, creators can choose from various modes such as collaborative creation (Copilot), swipe review (Swipe), or long‑term hosting (Autopilot), and only when Moras actually helps the influencer sell products and earn affiliate commissions does the system share revenue proportionally.

This strategy caused Moras’s user base to grow rapidly, with overseas registered influencers expanding from hundreds to nearly 100,000, and the cumulative GMV in the first quarter after official launch directly surpassed $10 million.

‘Creating real value is always more important than saving money; all users want to earn more,’ Wang Ming said plainly.

Besides helping people make money, the product also aims to capture local transactions within the vague intents of specific audiences.

A typical example is Karpo.

Since it can’t compete with big tech firms and hundred‑billion‑dollar unicorns in the downstream tasks of “helping people book certain tickets and canceling bills,” Karpo chooses to shrink the battlefield to the “evenings, weekends, and holidays” of a specific high‑spending demographic, moving one step forward to capture users’ vague intentions before they form explicit search terms.

The previous generation of the internet satisfied users’ explicit needs; users had to translate “I haven’t slept well lately” or “I want to relax this weekend” into concrete products themselves.

Karpo found that more than 70% of these user intents are extremely vague; they might just say “Food” or “Nature walk.” This gives Karpo an opportunity to intercept the intent.

To achieve this, Karpo bypasses the mass market and instead precisely targets urban youth aged 25‑35 in metropolitan areas like New York, San Francisco, and London—mid‑to‑high income, without childcare responsibilities, and passionate about exploration—using cross‑dimensional taste to turn vague emotions into orders.

For example, addressing the frequent pain point of users “want to go out to eat but can’t decide what,” the team developed a four‑quadrant visual slider called “VibePad”—users simply slide to choose whether they want tonight to be “relaxed and casual” or “formal and elaborate,” and the system instantly generates contextual prompts, delivering customized action plans ranging from street‑side eateries to Michelin‑starred restaurants, and automatically books the restaurant after credit‑card authorization.

By narrowing the proposition, Karpo has achieved high‑ticket‑price commercial monetization.

Currently, a typical individual user generates an average of $500 to $600 in real transaction volume per month through Karpo, with heavy users reaching as high as two to three thousand dollars.

By leveraging the purchasing authority of high‑net‑worth individuals, Karpo has also smoothly enabled commission‑based transactions with large chain merchants and ticketing platforms, as well as revenue sharing from linked‑card payments with financial institutions.

Is there a third way?

A group of hardware‑focused entrepreneurs in Silicon Valley propose the solution:to build a dedicated personal computer for the Personal Agent, using hardware to lock in the top 1% of professional users who are most willing to spend.

In early October this year, Oki, a team founded by post‑2000 Chinese‑American entrepreneur Luofei Chen from UC Berkeley and selected for the Y Combinator W25 incubator program, officially launched overseas a personal computer specifically designed for a 24/7 resident Personal Agent— Oki Home, priced at $1,799 for the initial release. Within just three days of launch, the first batch of pre‑order slots sold out completely.

The integrated hardware‑software Personal Agent has a clever business logic: it can generate profit by selling the hardware device directly, while also locking users’ contextual assets to the local hardware. Future opportunities include offering edge‑cloud collaboration, model and skill marketplaces, and other extensible services, greatly raising the commercial ceiling.

Wang Ming made a similar prediction in an interview with “Jiazi Light‑Year”:

“The current form of Personal Agent products is most likely a transitional stage, as they are mostly appearing on phones now. But a Personal Agent needs to continuously gather user context to understand you deeply, so it is very likely to be a seamless blend of software and hardware—perhaps in the future everyone having an AI glasses isn’t impossible.”

Obviously, there is no unified standard answer for Personal Agents yet.

But whether in the US or China, generic or vertical, consumer or enterprise, software or hardware, the war for Personal Agents will only just begin in the autumn of 2026.

(Cover image source: AI generated)

END.