Meta's agent Muse makes personal agents popular; retention and ecosystem barriers remain unresolved

In the month and a half since launch, Meta's agent Muse has shown staggering user growth. The latest data from Sensor Tower shows that since its September launch, Muse has accumulated 10.9 million downloads, of which 10.2 million were in the US.

In Meta's just-released Q3 earnings report, it said that about 25% of Muse users use it daily, more than 25% of users shop through Muse, and the average order value exceeds $200. Meta closed higher on the day the earnings report was released.

Muse's popularity has once again raised market attention to personal AI agents, and major domestic companies are also densely laying out their plans. According to media reports, ByteDance is developing a personal assistant called "Xiaodou"; the Tencent Cloud Lightweight Cloud team has launched a cloud-based Agent hosting product LightVela; Qwen also announced its plan to build a Personal Agent at the Yunqi Conference in September.

In addition to major companies at home and abroad, entrepreneurs are also eager to get involved. On October 21, Today AI, which focuses on AI personal assistants, completed a new round of financing worth $200 million, with investors including Tencent, Sequoia, IDG, Ant, Transsion and other institutions.

After chat, coding and office work, the wind of personal Agents is blowing again. How is it different from AI's previous functions, and what do manufacturers need to do to turn personal assistants into a good business?

At Meta's Q3 earnings call recently, Muse became the absolute focus of attention. Zuckerberg mentioned Muse more than fifty times in his opening remarks, while Facebook and Instagram were mentioned fewer than 20 times in total.

Payment service provider Plaid told Meta that it had never seen such high activity in any other service before. In the first week after launch, Meta even had to ask online payment company Stripe to urgently raise its credit limit because transaction volume exceeded everyone's expectations.

Later, someone saved $1,250 on a car purchase with Muse, someone canceled a book subscription his wife had forgotten to cancel, and someone dug out a $200 gift card from deep within an account.

This kind of extra spending in daily life is essentially a "hassle tax": people get lost in a flood of complex information, making it hard to find the most valuable information and make the most efficient decisions. This is where AI has room to create value, just as it does in enterprises. AI can reorganize your personal life and take charge of things you have no energy for or do not want to manage, giving you a more convenient and efficient life.

But the value of personal agents goes far beyond that. It can also be a "personal growth coach" or "life recorder," and even become an important tool for personal entrepreneurship and creativity.

Today AI, released on the same day as Muse, is the same. After Baobian registered for Today and authorized travel and email information to it, Today AI generated a relatively accurate personal summary and some daily management suggestions.

This way of using AI is very natural. In fact, when AI companion toys became popular, they carried this expectation from users. According to Baobian, at some domestic exhibitions earlier, elderly people and corporate managers asked manufacturers whether companion toys could take on the responsibilities of a personal assistant.

But companion toys do not have access to payments, calendars and email. A real personal assistant needs to live in the digital world and operate real services on behalf of users. This is also why major companies moved first: they happen to hold real-world interfaces such as account systems, payment channels and content ecosystems.

Muse's emergence is not the first time Meta has set its sights on agents. On December 29, 2025, Meta announced the acquisition of Manus, which develops agents. About four months later, after a foreign investment security review, the National Development and Reform Commission prohibited the acquisition and required the transaction to be unwound. In August 2026, Manus announced the resumption of independent operations. Meta subsequently launched its self-developed personal agent Muse in September.

From the OpenClaw craze at the beginning of the year to the autumn of 2026, agents have gradually penetrated from work scenarios into life scenarios. When acquired by Meta, Manus mainly met the needs of work scenarios. But now, from Meta and Tencent to Today AI and Manus, everyone is looking at personal life/work assistants.

Three weeks after Muse was released, OpenAI also released its own personal Agent Dots, focusing on all-weather task execution and integration into workflows.

In this exploration, on one hand major companies develop personal assistants themselves, and on the other hand there is investment involving personal Agents. In addition to Today AI, another financing star is Manus, which completed more than $500 million in financing after becoming independent, also backed by Tencent.

OpenClaw fades, butlers take the stage: the threshold is lower, but the walls remain

The major companies' persistent layout in personal agents itself also reflects a certain current state of AI applications. Among the main battlegrounds for AI applications in 2026, coding was the first to be proven, with the strongest willingness among users to pay, but limited by usage scenarios, its user scale has a natural ceiling.

JetBrains' 2026 Developer Ecosystem Survey shows that 90% of professional developers already use AI coding Agents at work every week, and 68% use them every day. AI coding tools have become the daily working method of mainstream developers.

Office work is currently the most competitive, with Tencent WorkBuddy, Alibaba Qwen Office, ByteDance Doubao Work and Kimi Work rushing in, and free quotas continuously increasing. AI applications need new usage scenarios to prove their value.

From the perspective of AI capabilities, the time is ripe for personal assistants. The core features of personal assistants lie in proactive service and emphasis on user memory. The agentic capabilities of mainstream models can meet the needs of various tool calls, and AI memory and context understanding are also sufficient for daily C-end interactions.

Beyond capability, the barrier to use is a problem that needs to be solved. Although today's Agents, compared with the early OpenClaw, no longer require a series of cumbersome configuration steps, Agents are mainly suited to work scenarios such as programming and office work, and interaction remains a "problem-solving" loop: the user issues a task—the Agent completes the task—the user gives feedback—the Agent revises.

If an Agent is to step into managing daily life, previous products required finding the right Skill, building the right program/dashboard, and repeatedly conversing with the AI to adjust it. A stable, considerate personal Agent with plug-and-play functionality is far friendlier.

Personal Agents also have to face an old problem: the many barriers of internet products.

Muse's growth stems in part from the highly fragmented and cumbersome nature of US internet services. In China, where the mobile internet is controlled by big tech companies and is more convenient, third-party personal Agents may struggle to adapt.

The Doubao phone, which has features similar to a personal Agent, ran into difficulties: 155 apps, including WeChat and Taobao, refused to allow it access. Even Muse faces this barrier: Amazon banned Muse from calling its shopping platform's permissions.

In China, companies have already laid out clear "ecosystem plays." Besides ByteDance's Xiaodou, which is backed by the Douyin/WeChat ecosystem, and Tencent's LightVela, Ant Group's Lingguang also began in its October new version to support calling apps within its ecosystem such as DingTalk, to handle tasks that require continuous follow-up.

It can be expected that domestic personal Agents will most likely form several walled cities centered on big tech companies.

For startups that want to build personal agents, they will probably have to become part of a big tech ecosystem. While Today AI accepted investment from Ant Group and Tencent, it also turned WeChat and DingTalk into Agent entry points through its Channels feature.

Three unsolved problems: user stickiness, monetization and security

Beyond long-term ecosystem building, the personal assistant business also needs to prove itself with short-term data, such as retention.

Sensor Tower data shows that although Muse's next-day retention is as high as 73.3%, its 7th-day retention drops rapidly to just over 30%, and by day 14 Muse is left with only 22.5%.

Even so, Muse is still above the average 14-day retention rate of under 20% for similar apps. The overall low retention also shows that personal assistants have not yet cultivated high-frequency demand among users.

Another problem is making money.

Payment mainly provides more usage: the Power tier costs $20 per month for 500 million Muse tokens per week, and the Maximum tier costs $100 per month for 3 billion tokens per week.

Zuckerberg said at the earnings call that Meta will eventually charge a "small fee" on transactions facilitated by Muse, and he expects this fee to be borne by businesses.

He also said he plans to increase Muse's computing efficiency tenfold within a year, so that it can achieve profitability once revenue scales up.

Domestic products are still in the free customer-acquisition stage, with business models undetermined, but they are also likely to rely mainly on transactions, usage or advertising for revenue.

More importantly, Agents entering daily life may bring security risks. macOS security expert Patrick Wardle disclosed a zero-day vulnerability in the Meta Muse macOS app. Attackers could exploit the AI's broad permissions to hijack users' devices and transcribe sensitive data to their own servers. He successfully used Muse to take photos and write malicious files to disk, and in many cases users were unaware of these operations.

Meta subsequently released a hotfix patch, but argued that this was a local privilege escalation vulnerability that required an attacker to already be running malicious code on the user's machine, so the actual risk was low.

At the same time, after Muse launched, some users assigned tasks such as calling to buy insurance or book tickets to the Agent, but because the other party hung up as soon as they heard it was an AI, Meta once proposed a laughable "human concierge" solution, in which the Agent hands the user's calling task to Meta's human customer service, meaning the user's privacy would be known to a third party.

Personal Agents may be the closest AI has come to ordinary people's lives. What makes them unique is that we entrust our work and lives to a steward with no physical form that learns as it goes. How to make everyone feel safe entrusting it with these tasks is a test for every personal Agent.

This article comes from the WeChat public account "Baobian" (ID: baobiannews), author: Zhang Jingwei, published by 36Kr with authorization.