Why Muse Topped the App Store While In-App AI Assistants Flounder

On September 8, Meta launched Muse, a personal AI assistant that provides each user with an independent cloud-based virtual machine (Secure VM) complete with its own browser, storage, and compute. Within 10 days it topped the US App Store free chart, surpassing ChatGPT; by day 12 third-party estimates put global installs at 2.8 million, and Meta's stock surged 11.43% on September 21, adding nearly $200 billion in market value.

Muse Actually Gets the Job Done

Muse's Secure VM runs a Sentinel Agent that governs internet access and sensitive operations. Before sending emails or making purchases, it pauses for user confirmation. Payments use a one-time virtual card generated via Stripe Link, bound to a specific merchant and amount, expiring shortly after use. Even if prompt injection occurs, the attacker only obtains a soon-to-expire disposable card number. This combination of broad permissions and tight risk containment represents serious engineering effort.

Users delegate tasks such as planning a concert trip (searching flights, comparing hotels, adjusting for price changes or availability) or turning saved Instagram recipe videos into shopping lists. The results are concrete: money saved on insurance, cheaper flights and hotels, completed bookings.

In-App Assistants: High-Friction Chat in Low-Latency Scenarios

Contrast this with Taobao's "Wenwen", Meituan's "Xiaomei", and Alipay's "Zhixiaobao". Their entry points are prominent but usage is minimal. When asked to assemble a summer outfit for a slightly overweight man, the assistant returns a paragraph of text plus three product cards with links. The user must then manually inspect each card.

Human visual processing scans a waterfall feed of 20+ items in 200 milliseconds; a chat interface forces the user to type, wait for token-by-token generation, load cards, and iterate. A 30-second browsing task becomes a 3-minute conversation with lower information density. For high-frequency, low-latency scenarios like food delivery (open, sort by rating/distance, reorder, pay in 30 seconds via muscle memory), inserting a chat layer breaks the flow and feels like an interruption.

Agents Belong in Low-Frequency, High-Complexity Workflows

Users willingly hand off tasks that span multiple apps, involve tedious steps, and contain massive information noise: disputing with merchants, organizing invoices for reimbursement, comparing flights and hotels across five or six apps. These are low-frequency, high-complexity scenarios — the true sweet spot for agents.

Real-world needs never stay inside a single app. A weekend camping trip requires weather (weather app), gear (Taobao), campsite reviews (Xiaohongshu), coordination (WeChat), navigation (Gaode). Each in-app assistant sees only a slice; none understands the whole plan. Muse handles the entire cross-app workflow as a single unit.

From Geek Tool to Mass-Market Product

OpenClaw previously demonstrated this capability by letting AI operate a real computer environment (browser, files, terminal) for continuous task execution, but it required self-hosted infrastructure, model configuration, and security management. Muse packages the same idea into a safe, reliable, easy-to-use product scalable to billions. Early users noted similarities in workspace layouts and configuration files (including SOUL.md); Meta's Nat Friedman acknowledged OpenClaw's influence while stating Muse was built from scratch.

The Structural Conflict: Who Does the Agent Serve?

An agent, by definition, acts on behalf of a principal. Muse's principal is the user: its KPI is saving the user money and time — cross-platform price comparison, coupon hunting, form filling, price-drop refunds.

In-app assistants face an inherent conflict. Taobao earns via bid-based ads and transaction commissions; Meituan sells merchant placements. A truly user-centric assistant that surfaces the best value regardless of ad spend would undermine the platform's revenue base. Consequently, recommendations are filtered through ad systems, mixing bid-weighted results. Users detect this after a few uses and abandon the assistant — it becomes a slower search box that requires typing and waiting.

This dynamic played out publicly on September 20: Amazon blocked Muse from browsing and checking out on its platform, while Shopify CEO Tobias Lütke announced support for Muse checkout on Shopify stores. The ad-and-commission model resists user-aligned agents; the merchant-tool model welcomes them.

Platform Rules Cage Model Intelligence

Mobile-era product methodology focused on funnels, retention, and ecosystem moats — each app a walled garden of artificial rules (viral mechanics, coin systems, cross-store discounts). When LLMs arrived, teams instinctively stuffed them into existing rule sets: acting as shopping guides, explaining promotions, keeping users inside the app longer. The model's general intelligence is trimmed into a business-specific customer-service bot.

Muse takes a different approach: underneath every digital service lies the same primitives — DOM trees, APIs, input fields, click events, state machines. A truly general agent needs only to perceive interfaces, click buttons, and reason about next steps, working on any site, even never-before-seen ones. When an app redesigns, the general agent adapts by observing the new UI; hard-coded rule chains break overnight and require rewrites. This generality provides antifragility.

Good Agents Know When to Stay Silent

A good agent works silently in the background: price comparisons run quietly, price-drop refunds are filed automatically, and only the final result is presented. Its goal is task completion. In-app assistants, however, are optimized for dwell time — they need the user's attention in the chat window as part of their KPI. The user wants results; the assistant wants process. These objectives are fundamentally opposed.

Industry Response and Remaining Hurdles

Muse's breakout has ignited the sector: OpenAI is preparing a comparable personal assistant, Manus launched the standalone Cue app, Grok Bot and Instinct pursue the virtual-machine direction, and ByteDance is reportedly testing a personal assistant product. "Personal Agent" has become the new focal point.

Yet three gates remain: Permissions — how much access will users grant? Security — real browsers, accounts, and payments mean risk can never be zero (Meta itself acknowledges this). Access — will third-party platforms allow the agent in? Amazon and Shopify demonstrated opposite answers on the same day.

Success depends not only on model intelligence but on how many permissions the agent can obtain, how many apps it can enter, and how much trust users place in it.

Conclusion

The gap between Muse and in-app assistants is not model capability — it is stance, delivery, scenario selection, and the understanding of what "agent" means. The first character of 代理 (agent) is 替 (act on behalf of). An assistant that cannot break its platform's commercial loop and unconditionally favor the user remains a glorified search bar with a chatbot skin. Users have voted with their feet: they want the thing done — a booked itinerary, a confirmed ticket. Three product cards and a block of copy? Nobody wants that.