By Wang Huiying

Edited by Ziye

In September 2026, the global AI community was abuzz with one term—personal Agent.

In just thirty days, nearly all major players at home and abroad simultaneously launched their own personal Agent products.

Meta's Muse surpassed 2.5 million downloads within thirteen days of its release, temporarily knocking ChatGPT off the top spot on the U.S. App Store's free chart. OpenAI introduced dots, proposing the concept of "always-on agents," while Manus launched a standalone app named Cue, equipping each Agent with an independent email, phone number, and digital wallet.

Across the ocean, China's AI scene was equally vibrant.

At Alibaba's Cloud Town Conference, QianWen confirmed its focus on Personal Agents, aiming to provide continuous services centered around individuals. Doubao had initiated its personal Agent project, codenamed Spell, as early as April this year. Following Muse's explosive popularity, the Doubao team accelerated the development of its personal Agent product. Tencent also hinted at the launch of a personal intelligent agent through Yuanbao.

This list spans China and the U.S., covering nearly every ecological niche from social media giants to cloud providers and startups. The density of activity rivals only that seen during the week of ChatGPT's release in AI's brief history.

The concept of Agents is nothing new.

2025 marked the breakout year for AI Agents, with tech enthusiasts scrambling for Manus invitation codes. In early 2026, OpenClaw sparked a nationwide craze for "AI shrimp farming," enabling Agents to operate computers 24/7 on behalf of humans.

Behind players' repeated bets on Agents lies an industry-wide ambition to end the era of AI chat dialogues and secure the primary entry point for the next generation of digital life. While both waves propelled Agents into the mainstream, most users remained unclear about what Agents could actually do.

This time, Agents have finally reached the tipping point for mass adoption. The focus has shifted away from model benchmarking or chasing better chatbots to competing for the next AI entry point.

After all, entry point competitions in any era offer only one window of opportunity.

1. Global Players Enter En Masse: The Personal Agent Race Begins in Earnest

After three years in the large model era, Meta finally delivered a blockbuster.

On September 8, an AI product named Muse officially launched, amassing over 2.5 million downloads in the U.S. and Canada within just 13 days. Its mobile daily active users reached 642,000—nearly triple ChatGPT's initial launch figures—and topped the U.S. App Store's free app chart for two consecutive weeks.

Capital markets reacted even more directly. On September 21, Meta's stock surged 11.43% in a single day, adding over $190 billion to its market value. By late September, Meta's stock had climbed more than 25% since the product's release.

This unannounced viral hit is Meta's personal Agent product, whose core breakthrough lies in shifting AI from conversational responses to task execution.

Source: Muse official website

Running on independent Secure VM cloud virtual machines, Muse eliminates the need for users to maintain conversation windows. By simply stating a goal, it autonomously crosses email, calendars, e-commerce, and social apps to advance full-chain tasks, only querying users at critical nodes like payments or permission confirmations.

From organizing emails and reminding users of expiring coupons to booking specialist doctors and comparing prices for trips, Muse continuously follows up in the background until completion.

The market's enthusiasm exceeded expectations, with Muse merely firing the starting gun for this global-scale explosion.

Within three weeks, the entire AI industry seemed to coordinate a mass Sprint , instantly thrusting the personal Agent track (arena) into full-scale competition.

On September 28, startup Manus launched its personal Agent app Cue, equipping each agent with an independent email, phone number, digital wallet, and even a dedicated cloud computer. At 12:00 AM on September 30, OpenAI formally released Dots at its developer conference. Each Dot operates in an independent cloud environment, accessing over 4,000 apps through ChatGPT's plugin ecosystem, enabling long-term background task tracking and proactive progress updates via SMS, Slack, and other channels.

Another personal Agent company, Instinct, rapidly gained capital attention. On September 28, Instinct announced a $1 billion financing round, valuing it at $10 billion. Its product helps users plan trips, purchase daily necessities, book tickets, and manage subscriptions via text or voice.

While overseas vendors unleashed full firepower, the true indicator of the track 's (arena's) heating up was the speed of domestic players' responses and product rollouts.

On September 22, Alibaba announced at its Cloud Town Conference an accelerated push for QianWen's Personal Agent, explicitly shifting from "users accessing QianWen" to "my QianWen." Zheng Sishou, QianWen's product lead, stated it would leverage the Agentic capabilities of the Qwen 3.8 series models to provide continuous, personalized intelligent services for 300 million users.

ByteDance responded even more directly. According to Sina Tech, after personal assistant products like Muse and Instinct gained overseas attention, Doubao accelerated the release of its personal Agent capabilities, planning a deeper integration with its Spell project. The new product, jointly developed by the Spell project and Doubao's conversation team, is expected to launch soon.

The Spell project, Doubao's internal test product for personal assistants since April this year, was led by the Doubao mobile assistant team but delayed for formal release due to factors like the Doubao mobile launch.

Tencent is also following suit. According to Zimu Bang, Yuanbao will introduce a personal intelligent agent, though its product form and launch timing remain unconfirmed.

Another product is LightVela, built by Tencent's Lightweight Cloud team. Launched in April this year, LightVela entered beta testing in May and public testing in August, positioning itself as a long-term online cloud AI companion.

Its official description as a "cloud Agent hosting product" allows users to configure dedicated intelligent agents in the cloud, operating 24/7 with support for long-term memory, custom models, and skills. Users can issue commands via WeChat, WhatsApp, Discord, Slack, and other platforms.

If accurate, Yuanbao resembles Muse, while LightVela competes with OpenAI's Dots—the former as a consumer-facing standalone Agent product, the latter as cloud-hosted Agent infrastructure.

Manus, which resumed independent operations less than a month ago, also launched its personal Agent app Cue on September 28, currently in early access. Cue's core design assigns each Agent an independent email, phone number, cloud computer, and digital wallet, enabling Agents to send messages, answer calls, and even autonomously complete payments within budget limits as "themselves."

Reviewing this wave of dense ( dense =intense) releases, all vendors agree: personal Agents are highly likely to become the next AI entry point after search engines and app stores. Whoever first captures user mindshare will control the next era's traffic initiative.

2. Securing the Next AI Entry Point: The Three Waves of Agent Evolution

September's dense ( dense =intense) launches made one fact clear: personal Agents are becoming the fiercest entry point battle in the large model era.

This war's eruption stems not from a sudden technological breakthrough but from a fundamental shift in Agent product design.

To understand this shift, consider a scenario.

American entrepreneur Joe Devoy uploaded his car insurance policy and requested a cheaper option without altering coverage. Within five minutes, Muse not only found a lower quote with the same coverage but also autonomously completed the new policy enrollment and old policy cancellation, saving him $3,500 annually. He only needed to confirm payment at the end.

Source: Muse official website

Another user tested daily shopping: asking Muse to find a green down jacket meeting specific requirements. Instead of merely returning product links, Muse automatically searched for real-time discount codes across platforms, applied the best offer, and completed the purchase at the lowest available price.

Similar scenarios emerged domestically. Nubia NaviX Ultra users tasked Xiaodou with multi-step requests like "book high-speed rail tickets + sync to DingTalk calendar + send reminders to travel companions," achieving an 87% success rate—up nearly 40 percentage points from six months prior.

These scenarios share a common thread: personal Agents have fundamentally altered human-machine collaboration. Users no longer need to break down tasks or monitor progress; they simply state an objective, and the Agent handles execution autonomously.

In contrast, 2025's Agents operated entirely differently. That year, tool invocation became standard for major models, with Manus's popularity demonstrating that large models could call external APIs and decompose multi-step tasks. However, these Agents were essentially functional modules attached to chat dialogues, still operating under a "user says one thing, AI does one thing" response model.

That wave validated the feasibility of integrating large models with tools but remained trapped in dialogue-driven frameworks. Users perceived them as smarter chatbots, not as assistants capable of independent work.

The 2026 "lobster craze" took a bigger step. OpenClaw demonstrated AI directly controlling computers—moving mice, opening browsers, reading/writing files, and sending emails. Theoretically, a single "lobster" could perform all human computer operations.

Developers rushed to "raise lobsters," using them to organize reports, auto-reply to emails, or even automate workflows. Yet few ordinary users persisted. Deployment required environment configuration and parameter tuning, creating high entry barriers.

Stability issues plagued many "lobster" products, which often forgot task objectives mid-process, clicked wrong buttons, deleted files, or repeated operations endlessly—forcing users to clean up messes.

While the "lobster craze" validated AI's potential for desktop control, it remained confined to niche developers' toolboxes, far from mainstream users.

The essential difference in this new wave of personal Agents lies in redefining products from tools to substitutes, with Agents transforming from dialogue instruments to execution entities.

Users perceive this shift directly. A Cue tester reported assigning an Agent to collect data and organize it into Excel on their computer. After closing the software, the Agent continued working in the cloud, with the file ready upon their return.

These subtle changes mark the core divergence between this wave and its predecessors. In earlier iterations, humans dominated tasks while AI executed instructions, requiring constant guidance and error correction. Now, humans merely define objectives and boundaries, while AI plans paths, executes steps, handles exceptions, and advances progress—truly freeing users from process management.

This fundamental change has made personal Agents the industry's consensus for the next AI entry point.

Dialogue-based products in the large model era are inherently transient, with no user retention or stickiness. Once model capabilities homogenize, forming genuine barriers becomes difficult. Personal Agents differ: they remember your work habits, lifestyle preferences, social relationships, and routine processes. The longer you use them, the better they understand you. As they accumulate more of your workflows, life patterns, and social connections, switching products becomes increasingly costly.

This data-and-habit-driven locking effect surpasses any single app. Once usage inertia forms, it becomes the default entry point for users' digital lives.

This explains why all vendors are investing heavily at this juncture. Entry point competitions offer only one window, and whoever cultivates user habits of "delegating tasks to Agents" first will dominate the next era's traffic narrative.

3. The Marathon of Personal Agents: Commercialization and Ecosystems Remain Decisive

Since entering the large model era, the tech circle has seen wave after wave of hype—each with eruptions and lulls, bets and exits. As all players converge on personal Agents as the next AI core, deeper competition begins.

Personal Agents aren't a track (arena) where a single launch determines victory. They demand sustained investment in computing power, model iteration, scenario expansion, and security maintenance—all requiring substantial resources.

To avoid becoming a short-lived trend, success hinges on two pillars: ecosystem and commercialization.

Judging by usage scenarios, a personal Agent's capability ceiling depends not on the large model's parameters but on its ability to integrate applications, cover scenarios, and connect data.

Conversational AI competes on response quality, with users switching freely. Personal Agents compete on task execution, with capabilities rooted in ecosystems. The more services they access, the more tasks they complete, and the stronger user dependency becomes.

This dependency transcends mere habit inertia; it involves comprehensive sedimentation of workflows, life patterns, and social connections. In this sense, ecosystem competition isn't about API integration quantity but about binding depth to users' digital lives.

An Agent's ability to handle tasks for users depends on how many "doors" it can open—email, calendars, payments, shopping, travel, social media—with each door representing an independent application ecosystem.

Domestic vendors hold natural advantages and structural challenges in this dimension.

Their strength lies in deep understanding of localized scenarios.

QianWen's open platform now covers over 20 fields, including local life, logistics, job hunting, sports, and gaming, with over 1,000 partner applications and integrations with services like SF Express, Tian Goose home (Tian'e Daojia), Hello Chuxing, and Ziru Rentals.

ByteDance integrates Doubao, Feishu, and Volcano Engine, with Doubao Work and Feishu sharing account systems and richer contexts. Agents can understand users' work rhythms from Feishu's chat records, documents, and calendars.

Tencent's unique strength lies in the WeChat ecosystem, whose billion-plus monthly active users form the largest distribution channel. Once users' work communications, payment records, and life services fully sediment in WeChat, the Agent's understanding depth becomes unmatched by standalone apps.

Yet these advantages come with barriers. Most domestic core services remain locked within individual app silos. An AI "butler" capable of booking flights, ordering food, and replying to emails requires not just cloud computing power but also keys to every digital "door."

The issues encountered by Doubao Mobile Assistant on Nubia's NaviX Ultra serve as a microcosm. Despite the phone being mass-produced and available for sale, automation operations within super apps like WeChat and Taobao are restricted. The most anticipated capabilities of AI are hindered by the willingness of ecosystem partners to collaborate.

Image source: Nubia Mobile's WeChat official account

In contrast, Meta's Muse operates on cloud-based virtual machines, unrestricted by mobile phone systems or super apps. For domestic manufacturers to develop Agents within someone else's operating system and app ecosystem, the difficulty is on a completely different level.

If the ecosystem determines how fast a product can run, then commercialization determines how long it can sustain. The previous two waves of Agent failed to reach the mass market, primarily due to immature product forms and a lack of clear, sustainable business models.

After the initial technological hype fades, without revenue to support ongoing R&D and operations, the wave naturally recedes. However, a notable difference in this wave of personal Agents is that mainstream players have designed their commercialization paths from the outset. Instead of developing products first and then seeking profitability, they are advancing product implementation with clear business logic.

Muse employs a combined model of subscription and transaction commissions, while OpenAI Dots relies on a revenue-sharing model through its plugin ecosystem. However, similar to ecosystem challenges, overseas commercialization models are difficult to directly replicate in the Chinese market.

Although domestic players have not yet disclosed specific commercialization models, referencing the pricing systems of previous conversational products, the commercialization paths for personal Agents are essentially extensions and upgrades of existing models. Three primary paths have emerged domestically.

ByteDance may extend a tiered subscription model. Building on Doubao's established three-tier subscription for personal users and a SaaS pricing framework based on seats and usage for enterprises, personal Agents are likely to be integrated as advanced capabilities into the existing membership system or offered through dedicated task quota packages.

Alibaba's Qianwen may continue its dual-track model of general subscription plus scenario-based value-added services. Basic transactional capabilities are covered under membership subscriptions to offset costs, while transactional scenarios like shopping, travel, and local services leverage Alibaba's ecosystem to earn commissions based on completed transactions, unlocking incremental revenue streams.

Tencent may also follow the traffic logic of the WeChat ecosystem. Basic Agent functions are embedded for free in mass-market scenarios to achieve rapid penetration, while cross-app advanced processing and exclusive services are monetized through membership enhancements. For B2B, it relies on WeChat Work's seat-based charging model, leveraging users to amortize customer acquisition costs.

Regardless of the chosen path, the entire industry faces two common challenges: First, the entrenched habit of C-end users expecting free services makes cultivating willingness to pay a lengthy process. Second, the high cost of integrating data and permissions across apps and systems not only restricts the Agent's capabilities but also affects users' perceived value of paid services.

Ultimately, the competition for personal Agents is never a 100-meter sprint but a marathon spanning several years. The spotlight at product launches is just the starting point, and iterative product features are merely the foundation. The true outcome will be determined by the depth of the ecosystem and the health of commercialization.

The ecosystem is the foundation—it determines how much value an Agent can create for users and how deeply it can engage them. Commercialization is the lifeblood—it determines whether manufacturers can sustain R&D investment, optimize user experience, and maintain endurance in long-term competition. Both are indispensable.

For the entire AI industry, this wave of personal Agents marks a turning point where generative AI transitions from the era of conversational interfaces to intelligent action. Only those who can establish a robust ecosystem and a viable commercialization path will ultimately secure their place in the next generation of AI entry points.