CICC: 2C Agents are moving from dialogue to task execution; Meta Muse's core advantage lies in its distribution network of 3.6 billion users and its trust infrastructure.

AI agents are moving from conversational tools to task execution. Meta's launch of the personal AI agent Muse has also shifted the focus of competition in 2C agents from model capabilities to connecting, executing, and continuously following up on real-world transactions.

Compared to previous Agent products, Muse's core change lies in its further integration with real-world transactions . CICC Research points out that Muse possesses capabilities such as asynchronous background execution and long-term memory, connection to real accounts, isolated execution, and approval gating. Ideas and Feed demonstrate the rudiments of a recommendation-based Agent. Meta also boasts a distribution base of 3.6 billion daily active users, as well as long-term content and behavioral context accumulated from WhatsApp, Instagram, and Facebook.

In other words, Muse's core advantage lies not only in its AI capabilities, but also in Meta's existing user distribution and trust infrastructure, which is something that other agent products find difficult to quickly replicate.

CICC Research believes that 2C Agents represent a significant evolutionary direction for AI applications following dialogue, with the key issue not being the direction itself, but rather the timing. As model and agent capabilities improve and token costs decrease, technological and cost conditions continue to improve. However, uncertainties remain regarding user expectations for task effectiveness, error tolerance, and migration costs. The speed at which 2C Agents scale will be determined when both technological effectiveness and usage costs simultaneously surpass user thresholds.

On September 8th, Meta officially launched Muse in the US, targeting users aged 18 and older, across iOS, Android, web, and WhatsApp. Powered by Muse Spark, Muse breaks down long-term goals into action plans and executes them in the background, covering scenarios such as web operations, emails, calendars, forms, travel bookings, and shopping. On its second day of release, Muse briefly rose to number 3 on the US App Store's free app chart and also ranked in the top three of the productivity charts.

Meanwhile, startups are also accelerating their entry into the market. Instinct can connect to email, messaging, and calendar via SMS and phone calls to complete tasks such as travel bookings, restaurant reservations, shopping, subscription cancellations, medical appointments, and email processing. From meta-level platforms to startups, the competition among 2C agents is further shifting from "what can be answered" to "what can be done for the user."

Muse's four capabilities: from single interaction to continuous execution

First, background asynchronous execution and long-term memory.

Muse can retain user preferences and task information across conversations and continue advancing predetermined goals even after the user closes the application. Goals are used to organize long-term goals and track task progress, while Artifacts store execution results as schedules, documents, or continuously updated dashboards, enabling tasks to be continuously tracked and accumulated.

CICC Research points out that compared to products that require users to launch a browser to execute tasks, Muse can continue working in the background according to a plan or related events. This gives the Agent a certain degree of continuous operation capability, and tasks are no longer limited to one-off, instantaneous responses.

Second, connect real accounts with external services.

Muse can be integrated into various scenarios such as email, calendar, payment, health, smart home, and e-commerce, and connects to services like OpenTable, Ticketmaster, and Spotify. Its connection methods primarily fall into three categories: built-in connectors, using user-provided credentials to call public APIs, and using browser operations to override services lacking available interfaces.

Account connectivity enables Muse to access existing user information and services, linking information acquisition, request communication, and subsequent operations. This reduces the need for users to repeatedly provide data, switch platforms, and perform manual operations, providing a foundation for cross-service tasks.

Third, implement isolation and approval gate control.

As agents begin to access sensitive information such as emails and payments, security and permissions have become crucial aspects of product design. Meta provides each user with an independent Muse Secure VM, with Sentinel acting as an agent to handle external actions; Meta also plans to launch a Confidential VM that supports user-owned keys by the end of the year.

CICC Research believes that this architecture, through its independent operating environment, credential protection, and action approval, establishes authorization boundaries, which helps limit unauthorized operations when the model is misled and reduces the burden of continuous user oversight. However, Meta also disclosed that its technical architecture still retains the ability to access virtual machines, so this mechanism does not mean that the platform and execution environment are completely isolated.

Fourth, Ideas and Feed explore recommendation agents.

In addition to Chat, Goals, and Library, Muse also features two proactive information feeds: Ideas and Feed. Ideas can suggest tasks based on the user's past conversations, calendar events, and connected accounts; while Feed generates a customized information stream based on the user-defined themes and pace.

CICC Research believes this indicates that agents are beginning to shift from "waiting for user questions" to "proactively identifying needs," representing an early sign of the evolution from search-based interaction to recommendation-based services. However, Ideas currently still relies on user authorization and confirmation, limiting its proactive nature.

Meta's advantages: Distribution, context, and execution foundation

Meta's key foundation in 2C Agents comes primarily from its existing user ecosystem.

According to the company announcement, the Meta app family reached 3.6 billion daily active users in June 2026. Users can directly use Muse through WhatsApp conversations without needing to rebuild their usage habits for a separate app. With user authorization, long-term content and behavioral information accumulated on Instagram and Facebook can also supplement personal context beyond chat logs.

CICC Research believes that WhatsApp can serve as the entry point for interaction and distribution, Instagram and Facebook provide personal context, and third-party connectors and payment infrastructure expand the scope of task execution. For Meta, this ecosystem combination constitutes an important product foundation for Muse.

However, Meta's ecosystem also faces practical constraints. CICC Research points out that compared to ecosystems like WeChat, which have integrated accounts, payments, and lifestyle services within their platforms, overseas service entry points are relatively fragmented. Muse needs to combine APIs and browser operations to cover more long-tail services. This model has relatively high connection and maintenance costs, and website redesigns may also affect operational efficiency and reliability.

In terms of business model, Muse offers a free tier and two monthly subscription options: $20 and $100. Power and Maximum subscriptions are $20 and $100 per month, respectively. CICC Research believes that in addition to subscription revenue, with the integration of consumer tasks such as shopping and booking, Muse has the potential to explore transaction commissions or revenue sharing in the future, but this model remains a potential path at present.

From "Search" to "Recommendation": Timing Remains Key for 2C Agents

CICC Research believes that 2C agents may undergo a model shift from "search" to "recommendation" in the future.

Current agents are closer to search: users proactively define problems, and the agent is responsible for solving them. A longer-term model might resemble recommendation systems, where the agent proactively identifies needs, defines tasks, and drives execution based on sufficient user context. This model lowers the barrier for users to proactively submit requests, but it also places higher demands on technical capabilities, user authorization, and costs.

From the perspectives of technology and cost, CICC Research divides product innovation into four stages: exploration, transitional innovation, all-round innovation, and the strong getting stronger. Only when the technological effect exceeds the user acceptance threshold does a product have the foundation for commercialization; only when the cost is further reduced to an acceptable range can it more easily enter the stage of large-scale commercialization.

For 2C agents, technological productivity depends on factors such as model capabilities, agent capabilities, contextual understanding, memory, and application ecosystem, while costs are related to computing power, tokens, and R&D costs. Currently, it's relatively certain that technological capabilities are still improving and token costs are trending downwards; what's more difficult to judge is users' tolerance for errors in agents completing tasks such as booking, shopping decisions, and email processing, as well as the costs of migrating from existing solutions.

Based on this framework, CICC Research proposed two possible paths.

One scenario involves technological capabilities exceeding user demand thresholds, but cost reduction lagging behind. Agents might initially enter the market with technologically controllable and relatively cost-effective scenarios such as general office work, and then gradually expand to comprehensive tasks as technology advances.

Another scenario involves rapid technological advancements coupled with a quick decrease in costs, allowing 2C agents to directly enter a more comprehensive application phase. In this case, platform-based products with rich user context, application ecosystems, and distribution channels are likely to scale up more quickly.

Therefore, CICC Research believes that the time required to transition from the transitional innovation stage to the all-round innovation stage is an important variable for observing the future competitive landscape : a shorter interval may make it easier for internet platforms to leverage their existing ecosystem and channel advantages; a longer interval will give startups more time to build product capabilities around specific scenarios.

Risk warning and disclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.