Less than a month after its rebirth, Manus has equipped the agent in its new product with four essentials: an email account, phone number, wallet, and computer.

Cue, released on September 28, is a standalone personal agent application. Each agent can communicate with the outside world using its own email and phone number, make payments within a user-set budget, and continue tasks on its own computer. The company aims to handle communication, payments, and follow-ups after task delegation.

However, twenty days earlier, Manus's 'ex-partner' had already reached this point. On September 8, Meta launched Muse in the U.S., providing users with a dedicated cloud computer. The company stated that tasks could continue even after the app was closed. On September 29, OpenAI announced dots, allowing assistants to keep working between conversations and returning only when user input is needed.

If assistants can handle tasks like 'arrange my business trip' or 'buy fresh groceries for the week,' some decisions might be made before users even open shopping or travel apps. In the past, platforms competed for user visits; now, they must also vie to be chosen by users' assistants.

China's 'BAT' are also vying for this position. ByteDance boasts a vast AI app user base and experience in having assistants perform tasks within mobile apps. Tencent holds the social connections, merchants, and services within WeChat. Alibaba has Taobao, Gaode, Fliggy, and Alipay, covering everything from product discovery to payment, with many steps already within its ecosystem. While each company brings different strengths, their goal is the same: 'how to get users to delegate more tasks to a single assistant and ensure those tasks are truly completed.'

This is what makes personal agents more significant to platforms than chatbots. They could influence where the next order goes. But at the start of this competition, Amazon has already shut the door on Muse: just because users are willing to have assistants make purchases doesn't mean stores are willing to serve them. Who retains this entry point must pass two tests: whether users continue to delegate tasks and whether external services are willing to let them handle those tasks.

Manus and Meta: Learning New Lessons After Parting Ways

The two companies originally had the opportunity to collaborate on this.

Late last year, Manus announced its integration with Meta, expressing a desire to bring subscription services to Meta's enterprise and user base. On September 1 this year, Manus announced its return to independent operations. More than twenty days later, both have unveiled their own personal agent products. The user engagement and execution capabilities they once planned to combine through a single deal must now be developed separately.

Meta already has extensive experience interacting with people. Mark Zuckerberg stated at a July 29 earnings call that 3.6 billion people use at least one of its apps daily. Regarding personal agents, he emphasized that while engineers are willing to spend time configuring tools, mass-market products must be ready to use out of the box to be accepted by billions.

Muse, therefore, places the familiar chat interface front and center. Tasks can be delegated via a WhatsApp message. However, transitioning from 'send a message' to 'send an email on my behalf, make a payment on my behalf' requires another level of trust. Unsatisfactory recommendations can be dismissed with a swipe; a misdirected email, however, won't disappear simply by closing the app.

Meta has set different permission levels for these actions. According to the company's disclosed design, Muse will seek confirmation before sensitive operations like sending emails or making purchases, allowing users to adjust permissions or disconnect the service. The approach is to grant visibility first, then determine how much to delegate—a gradual authorization process.

A group of users has already begun trying Muse. Internal data reported by The Information shows that over 3 million users submit at least one instruction per week. Meta quickly identified a group with more intensive needs. On September 29, the company announced the expansion of Muse's small business features, integrating tools like stores, ledgers, and customer records.

Tom Mulholland, a grocery store owner in Iowa, says he works about 65 hours a week. While skilled at cutting steaks and making sausages, he still has to handle other roles in running a store. He hopes to save time for tasks that generate revenue.

For Meta, merchant relationships could extend from one-time promotions to daily operations: preparing content, organizing customer information, and following up on unanswered emails. These tasks are trivial yet frequent enough. If assistants can handle them, platforms gain another way to retain merchants.

Manus, on the other hand, has a different lesson to learn. Positioned as a general-purpose agent, it has built products around research, automation, and complex tasks. Xiao Hong's statement on social media, as relayed by media, compares Manus to selling computers: people buy computers mainly for work but also use them for watching videos and playing games. He hopes to retain the tool's versatility.

This update integrates general capabilities into two products. The desktop app upgrades to Studio for continued creation, while Cue enables ongoing task delegation, both using the same Manus infrastructure. Studio's video editor leaves users a timeline. After receiving a draft, they can change the soundtrack, incorporate their own footage, and have the system continue editing. The more time users spend here, the more useful the tool sometimes proves.

Cue faces a different usage pattern. If there are no schedule changes or client responses, the assistant doesn't need to call the user back. If every step requires re-explanation, inspection, and correction, the time saved by users is quickly spent managing the assistant.

Xin Lichang believes that office agents and personal assistants require different retention metrics. Workstations should measure whether users continue creating; personal assistants should gauge whether tasks delegated once are entrusted again. Frequent user interaction with the assistant sometimes indicates a lack of trust.

Once a form is submitted, the task can end. However, contacting clients requires remembering promises, tracking unanswered messages, and knowing when to follow up. Cue allows multiple agents to divide tasks, but if handovers still rely on humans to fill in information, users are simply managing an additional team.

Meta aims to have existing user relationships shoulder more responsibilities, while Manus seeks to transform one-time executions into long-term delegations. However, users won't spend half an hour managing an agent to save one hour of work.

'B' Navigates Through Apps, 'A' and 'T' Explore Different Paths

Domestic companies are already exploring several distinct paths.

ByteDance's efforts began before Muse's release. Doubao launched a personal assistant beta test in April, codenamed Spell, led by the Doubao mobile assistant team, with plans to advance jointly with the Doubao conversation team. On September 30, TechNode reported plans for an independent app tentatively named 'Xiaodou.'

These two teams' capabilities could complement each other: one understands requirements, the other operates within apps. For example, rescheduling a flight requires knowing acceptable times, reading the order, and comparing price differences. Sudden price hikes necessitate re-inquiry; identity verification pop-ups require human intervention. What the passenger ultimately receives must be a rescheduled ticket and a clear account.

ByteDance has already encountered such boundaries. On December 6 last year, the Doubao mobile assistant announced it had disabled the ability to operate banking and internet payment apps due to fund safety concerns. While an independent app could expand reach, its actual access to apps and the actions it can complete must still be clarified. Technologies accumulated for mobile assistants won't automatically grant access to all services.

Tencent is pursuing a path closer to service internals. Tencent Customer Service stated that WeChat is testing a native AI assistant, 'Xiaowei,' in a small scope, supporting text or voice operation of some native WeChat functions.

WeChat already hosts people, merchants, and service pages. Connecting demands to these services could reduce the need for users to find additional entry points or repeat requirements. However, after locating a merchant, checking inventory, submitting reservations, and canceling orders remain distinct actions. The extent of service provider openness directly determines what assistants can accomplish.

Tencent Lightweight Cloud Team's LightVela focuses on enabling continuous agent operation. In its official introduction, the team recalled the March OpenClaw Public welfare installation activity and attributed the starting point of cloud services to the need for sustained, stable services after tool installation.

Xiaowei could connect services within WeChat, while LightVela provides cloud operation conditions. Additionally, on September 30, reports emerged that Tencent is internally developing a personal agent product, Handy Bot, with plans for an independent app.

How Handy Bot, Xiaowei, and LightVela will divide labor and share resources remains incompletely publicly explained. Labeling them as 'racehorses' obscures the distinct problems each aims to solve: acquiring users, obtaining service permissions, and enabling continuous task operation.

Alibaba is also attempting to extend one-time services into ongoing personal delegations. On September 22, Qianwen announced at the Yunqi Conference its acceleration in building Personal Agents, disclosing progress in accessing user-authorized personal data like health and fitness, aiming to integrate personal assistants into tasks requiring ongoing follow-up.

The competitive landscape isn't so orderly. On October 8, Manus's parent company, Butterfly Effect, announced completing a new funding round exceeding $500 million, with existing investors like Tencent continuing to participate. Tencent is both developing its agent and investing in Manus. Between startups and giants, both product competition and capital relationships exist; investment doesn't equate to aligned business operations or permissions.

Each company brings legacy businesses into this new competition. Existing users help assistants receive demands sooner, while merchants and services shorten execution paths. However, users don't evaluate by department—whether flight tickets are rescheduled, appointments succeed, or costs are reasonable—are the visible results.

After Acquiring Users, Platform Approval Is Next

Less than two weeks after Muse's launch, Amazon stood in its way.

On September 20, GeekWire reported that users attempting to connect Muse to Amazon Shopping began seeing prompts: unauthorized AI agents continuing access violate platform terms. Amazon confirmed blocking Muse, stating it had previously requested Meta exclude shopping websites from its services.

Amazon objected to agents not identifying themselves and handling account credentials. Meta's disclosed design states credentials enter secure storage, with Muse itself unable to see passwords. However, secure storage doesn't directly address whether service providers consent to agent access.

This dispute at least shows that user authorization and platform access are two issues requiring separate solutions. After receiving 'buy this for me,' the assistant still needs to find willing stores.

Two days later, Instacart made a different choice. Chief Technology Officer Anirban Kundu announced in a company statement plans to integrate fresh grocery shopping into Muse. His reasoning was straightforward: shopping demands arise in various places, and Instacart wants to be present.

On July 16, the company had already announced integration with Google Search's AI Mode. Users could generate shopping carts from dinner plans and, after connecting accounts, check out via the Instacart app or website. Conversations could start externally, but orders are still processed by its system.

Instacart's willingness to integrate with Muse stems from its calculations. Shopping demands can emerge from others' assistants, but commodity (products), inventory, and delivery remain handled by its system. The assistant might understand 'what four people eat tonight,' but stores still need to inform it of available ingredients, prices, substitutes, and delivery times. Changing conversation windows doesn't eliminate these tasks.

For Instacart, if Muse drives new orders and enables shopping to begin from a conversation, the partnership holds value. However, as more demands enter through external entry points, how much direct customer interaction merchants retain becomes a negotiating point later.

From this perspective, Alibaba might be among the most positioned domestic companies to first establish a functioning personal agent transaction chain. Taobao handles products and merchants, Gaode connects locations and local services, Fliggy manages flight and hotel bookings, Damai covers events and ticketing, and Alipay handles payments. These businesses already address parts of daily needs.

This positions Alibaba to skip a step: first connect existing services to assistants, then identify tasks users are willing to delegate long-term. However, having comprehensive internal businesses also introduces new issues. When users ask Qianwen to recommend a restaurant, flight ticket, or product, how will it handle choices between group and external options? The more services integrated, the more the assistant must explain whose judgment it represents.

Wall Street has begun calculating potential earnings. On October 6, Reuters cited Citi's forecast that Muse's annual revenue could exceed $27 billion by 2030, with roughly $23 billion from transaction-related income and $4.5 billion from subscriptions. In other words, transactions might prove more valuable than subscriptions in this forecast.

This remains a projection based on future user growth, transaction scale, and commercialization capabilities. Early data disclosed by Citi shows Muse has accumulated over 6.6 million downloads since launch, with 1.8 million daily active users. There's still a long way to go for an annual $27 billion business.

However, this revenue structure highlights another issue with personal agents. If someone tells Muse they want Italian food tonight on a $200 budget and a restaurant nearby, the assistant can compare ratings, distances, and prices before making a reservation. But which restaurant it ultimately recommends could also influence platform transaction revenue. Is a restaurant recommended because it better fits user requirements or due to commercial partnerships between platforms?

Even within giants, businesses worth investing in must be selected. On August 13, Tencent management had already indicated in reports that due to WorkBuddy's growth, the company prioritized resources for that app while reducing priorities for other AI products. Management also stated that inference computing investment is tied to business returns, with adjustments made if returns fall short.

Previously, such issues mainly arose in search results, ad placements, and product recommendations. With personal agents, users might not even review candidate lists. The comparisons and Filter (filtering) once done by humans are now compressed into a single assistant choice.

This is one reason platforms' attitudes are diverging. Amazon is unwilling to cede shopping entry points to unauthorized external agents; Instacart welcomes Muse to drive new shopping demands while continuing to provide products, orders, and delivery. Alibaba has the opportunity to leverage its ecosystem to directly connect some transactions to Qianwen. Behind these three paths lie original platforms recalculating: how much new business agent integration brings versus how much direct user interaction it loses.

Independent Agents need to secure these collaborations one by one. Large corporations can initially leverage their own services but must also consider how to manage relationships with merchants both inside and outside their group. The extent of openness, who completes transactions, and who users should turn to when issues arise all require further negotiation. Personal Agents can shorten the path from demand to payment, but many companies already operate along this route. They won't automatically surrender their business just because users switch to a different chat window.

Written at the End

Alibaba has already seen the price of being late to compete for AI entry points. In November 2025, Alibaba entered the consumer AI market with its QianWen App. Subsequently, QianWen rapidly expanded its user base through product upgrades, ecosystem integration, and large-scale operations. However, QuestMobile data from June this year shows that Doubao has 382 million monthly active users, while QianWen has 167 million—still more than twice as few.

Once users become accustomed to opening a particular product, convincing them to switch isn't so easy.

Personal Agents may make this habit even harder to break. Chatbots need to remember what users have said, while personal assistants must also recall which tasks users have entrusted to them—who needs to be contacted in advance, what money cannot be spent, which changes require prior approval, and how to handle issues when they arise. These details may hold more value than a mere chat history.

If users continue to rotate between several products, re-explaining the context every few days and handling key steps themselves, there's still an opportunity for latecomers. Delivering one task reliably often makes it easier to secure the next commission than promising to manage an entire day.

This also subjects the users, models, and services accumulated by large corporations in the past to a new test. Meta has user relationships, Alibaba has transaction infrastructure, Tencent has services within WeChat, ByteDance has AI users and mobile operation capabilities, and Manus attempts to leverage its experience in executing general-purpose Agents to handle some complex tasks first.

Each company can skip some steps compared to others, but no single advantage is yet sufficient to directly determine the outcome. In September, companies began equipping personal assistants with computers, email, and wallets, attempting to enable continuous operation and even spending money on behalf of users.

The next step is to see whether users are willing to entrust the same task to it a second or third time—and whether external merchants are willing to continue fulfilling orders for it. Only then will this entry-point competition gain a clearer progress metric than download counts and release order.

*The featured image and illustrations in the text are sourced from the internet.