Meta’s Muse Is What Google’s Gemini and Apple’s Siri Should Be

The message arrives while you are rushing through an airport, your phone buried somewhere in your bag and your next meeting already appearing on your calendar. Your flight has been delayed by three hours, which means the hotel booking you made for tonight may no longer make sense, the restaurant reservation you had planned for dinner is now inconvenient and the person you are supposed to meet tomorrow morning has not yet been told that you will arrive late.

Instead of opening five applications, searching through emails, changing a reservation, sending a message and then rebuilding the rest of the itinerary yourself, you tell your AI assistant what happened and let it work out the consequences. It checks the flight, finds the hotel reservation, identifies the dinner booking, looks at tomorrow’s calendar, contacts the restaurant and prepares the necessary changes, stopping only when it reaches a decision that requires your approval.

That is the moment the artificial-intelligence industry has been promising for years, but it is also the point at which Meta’s Muse becomes more interesting than another chatbot with a larger context window or a more convincing voice. Meta introduced Muse in September as a personal AI agent designed to take action on a user’s behalf, rather than merely answer questions.

It can browse the web, fill in forms, send email, book travel, make purchases and continue working after the user closes the application, returning when it needs approval or when circumstances change. Meta has built the system around a dedicated virtual machine, its own browser and a separate security layer called Sentinel that controls what Muse can do outside its environment.

The distinction matters because the AI industry’s central competition is quietly moving away from the question of which company has the most capable model and toward a much more consequential question: which company can persuade consumers to delegate meaningful portions of their lives to software.

Google already has the information infrastructure to make Gemini an extraordinary personal assistant, while Apple controls the hardware, operating system and applications through which hundreds of millions of people conduct their daily lives. Both companies have moved aggressively toward agents. Yet Muse provides a particularly clear demonstration of what those systems need to become because Meta has defined the product around delegation from the beginning: the user describes an objective, Muse develops a plan, interacts with the necessary services and continues working until it reaches the limits of its authority.

This is a fundamentally different proposition from the first generation of digital assistants. Siri, launched by Apple in 2011, normalized the idea that people could speak to a computer rather than navigate menus, while Google’s assistants made conversational search increasingly ubiquitous. But the underlying interaction model remained surprisingly close to conventional computing: the user asked a question, received an answer and then performed the consequential work.

An assistant could tell you the cheapest flight, but you still had to book it; it could identify a restaurant, but you still had to make the reservation; it could summarize a long email thread, but you still had to decide what to write and send the reply.

The new generation of agents is attempting to eliminate that gap between information and execution. Google said in May that more than 900 million people across 230 countries and more than 70 languages were using Gemini each month, up from 400 million the previous year, and it has been adding features designed to make Gemini proactive rather than purely reactive.

Google has introduced Gemini agents capable of working continuously in the background and Search agents that monitor information across the web, while Gemini 3.5 has been explicitly designed around complex agentic workflows.

Google’s advantage is difficult to overstate. Gemini sits on top of an ecosystem containing Search, Gmail, Calendar, Maps, Chrome, YouTube, Android, Photos, Shopping and a vast amount of information about the world.

If an AI assistant is ultimately judged by how much context it can bring to a task, Google has spent decades assembling the raw material. A person asking, “Find me a hotel in London for Tuesday,” can eventually expect Gemini to know not only what hotels are available but what the user has previously booked, which neighbourhoods they prefer, when their flight arrives, what meetings appear on their calendar and which emails contain relevant travel information.

The problem is that having the information is not the same as turning it into a coherent assistant. Google’s products were built as separate services, each optimized around a particular activity, and Gemini is now being asked to become the intelligence connecting them.

The company’s own product announcements show how rapidly that transformation is happening: Search agents can operate in the background, while Gemini itself is being positioned as a persistent assistant capable of proactive help rather than simply a destination where users type questions.

Apple’s opportunity is equally significant but structurally different. Its advantage is not the world’s largest search index; it is the device sitting in a user’s hand, on a desk, in a car or on a wrist. Apple’s redesigned Siri AI is intended to understand personal context across messages, email, photographs and other information, recognize what is displayed on the screen and perform actions across applications.

Apple gives a revealing example: a user can ask Siri to find a restaurant recommendation mentioned by a friend in a message or retrieve a hotel confirmation number from an old email, and Siri can use onscreen awareness to understand a conversation and then act on it in another application.

In other words, Apple already has much of the architecture required to build the assistant that consumers have imagined since Siri first appeared. What it has historically lacked is the willingness, or perhaps the technical confidence, to allow Siri to cross the boundary from answering a user to independently completing a chain of actions.

That boundary is now becoming less defensible as AI models improve. If Siri can understand a text about a dinner, identify the restaurant, check the calendar, find an available table and prepare the reservation, requiring the user to manually open each application begins to look less like a safety feature and more like an artificial limitation.

Muse’s significance is that Meta has made this agency the centre of the product rather than an extension of an established assistant. Meta says users can give Muse a broad goal, after which it can coordinate time and resources, create a personalized plan and continue advancing the work independently.

For a relatively simple task, that might mean sending an email or booking travel; for a more complicated objective, Meta envisions Muse adjusting its actions as circumstances change. The company says the agent can even work on goals such as selling a car for a better price, lowering a bill or modifying a training plan as a person’s schedule changes.

That is closer to employing a digital assistant than using a search engine. The difference becomes particularly powerful when memory enters the equation. Meta says Muse can remember information that a user mentioned only once and later use it without being prompted again. A recipe saved from Instagram can become a grocery list; Muse can suggest a menu for a dinner party and remember guests’ dietary restrictions before preparing invitations.

These are mundane examples, but they reveal where personal AI could become genuinely useful: not by producing spectacular answers, but by maintaining the small pieces of context that people ordinarily carry around in their heads.

That is also where the three companies are approaching the same destination from different directions. Google has the strongest claim to understanding the external world, Apple has the strongest claim to understanding the user’s device and personal computing environment, and Meta has chosen to make conversation and delegation the primary interface. The competitive question is whether consumers care more about the assistant’s underlying intelligence or about how naturally it fits into the way they already communicate.

Meta is betting on the latter, and its decision to bring Muse to its AI glasses makes that bet considerably more consequential. At its September Connect event, Meta said Muse would come to its AI glasses, allowing users to address the agent hands-free and have it act on what they are looking at. A user could ask about a product on a shelf, a flyer on a wall or a long list of school supplies without first describing the visual information to the assistant. Meta also said Muse would support long voice conversations while continuing to work in the background.

That begins to resemble an operating system for the physical world. A person wearing AI glasses does not need to photograph a document, open an application and upload it to a chatbot; the assistant can see what the user sees, understand the context and potentially act. The implications extend well beyond convenience.

The device could identify a product, compare its price, check whether it fits an existing purchase plan, remember that the user is shopping for someone with a dietary restriction and then order it. The same system could become a navigation layer, a translation tool, a visual search engine and eventually a persistent personal agent that follows its user throughout the day.

But the more capable these systems become, the more the central problem shifts from intelligence to trust. A chatbot that produces an incorrect answer may waste several minutes. An agent that sends the wrong email, purchases the wrong item, changes a travel reservation or exposes private information can create consequences that are much harder to reverse.

Meta’s own researchers acknowledge this directly. In developing Muse, they said the company had effectively handed an AI system access to inboxes, calendars and a computer shell and allowed it to operate unattended, only to discover that the system did not always behave as expected. The resulting security architecture was built on the assumption that Muse would make mistakes and that external content could attempt to manipulate it through prompt injection.

Meta therefore built a substantial security system underneath the agent. Muse operates inside an isolated cloud computer, while Sentinel separately determines whether the agent can communicate with external services.

Credentials are stored separately from the core agent, and the system can require user approval before sensitive actions such as sending an email or making a purchase. Meta says users receive an audit trail of what Muse has done and can control which services it can access. The company is also developing a Confidential VM intended to prevent even Meta from accessing the user’s encrypted data.

Those safeguards illustrate an important point about the agent economy: the decisive technology may not be the language model itself. It may be the infrastructure surrounding the model.

An agent needs permissions, authentication, memory, browsers, connectors, payment systems, monitoring and mechanisms for determining when it should stop and ask a human. The model is the reasoning engine, but the product is the entire system that determines what happens when that reasoning meets the real world.

Apple has an obvious advantage here because privacy has been central to its product positioning for years. Its new Siri architecture is designed to combine personal context with systemwide capabilities while maintaining Apple’s existing privacy model.

Google’s advantage is scale and integration: Gemini can draw on an ecosystem that contains enormous quantities of structured information and services. Meta is attempting to solve the same problem through the dedicated virtual-machine architecture behind Muse and by making the agent’s permissions explicit and granular.

None of this means Muse has won. In fact, the most interesting thing about the product is that its weaknesses expose how difficult the category will be. Meta acknowledges that Muse can make mistakes, and its researchers describe prompt injection as a fundamental threat because agents necessarily encounter information that they do not control.

An email, webpage, document or advertisement can contain instructions designed to manipulate the model, and an agent that simultaneously has access to private data and the ability to communicate externally can turn a seemingly harmless piece of text into a security problem. Meta’s security team says it has used red-teaming, external researchers and a bug bounty program to address those risks, including rewards of up to $300,000 for valid security reports.

The other unresolved issue is whether people actually want this level of delegation. For years, consumers have been trained to treat AI output with caution. An assistant that says something incorrect can simply be ignored; an assistant that takes an incorrect action requires a much higher level of confidence.

This creates an unusual psychological barrier for Google, Apple and Meta because the value of the product increases precisely as the user’s direct involvement decreases. The perfect agent is therefore one that users barely notice, but achieving that invisibility requires extraordinary reliability.

This is why Muse’s most important idea is not its model, its browser or even its ability to operate in the background. It is the change in the unit of interaction.

The basic unit of conventional software is the application: open Gmail, open Maps, open a browser, open a shopping app. The basic unit of an agent is the objective: plan the trip, deal with the bill, organize the dinner, prepare the report. The user should not have to know which application contains the answer or which sequence of buttons produces the desired result.

Google is moving toward that world because Search itself is becoming agentic. Apple is moving toward it because Siri increasingly understands personal context and can operate across applications.

Meta is moving toward it through Muse, whose entire premise is that the user should be able to state an objective and let an AI system determine the steps required to accomplish it.

That makes the title of this race somewhat misleading. The question is not whether Muse is smarter than Gemini or Siri, because model benchmarks and feature comparisons will become obsolete almost as quickly as they are published.

The more consequential question is which company can create an assistant that understands a person’s intentions, remembers the context behind them, acts reliably across services, knows when it is about to make a consequential mistake and asks for permission without turning every useful task into a sequence of approval screens.

Muse has a head start in one important respect: Meta designed it around that problem from the outset. Google’s Gemini and Apple’s Siri have to transform products that were created for a different era of computing into agents that can operate across the user’s digital life.

Their enormous existing ecosystems are advantages, but they are also constraints because the assistant must reconcile years of applications, permissions and user expectations.

If Meta succeeds, the significance will extend well beyond another AI product launch. Muse could help establish the model for a new kind of consumer software in which people stop navigating applications and start delegating outcomes. That would make Google’s information advantage and Apple’s hardware advantage less decisive because the interface between the person and the digital world would increasingly be the agent itself.

The irony is that the technology companies spent the first decade of the smartphone era teaching people where to tap, swipe and search. The next decade may be about teaching them that they no longer need to do any of those things for themselves.

The assistant that ultimately wins may not be the one that produces the most impressive answer when asked a question, but the one that understands what the user is trying to accomplish, carries the task across the fragmented machinery of modern life and knows exactly when to stop before the human needs to take over. Meta’s Muse is significant because, more clearly than either Gemini or Siri, it is being built around that proposition from the beginning.

Stay ahead of the Stories shaping our world. Subscribe to Impact Newswire and join our 
WhatsApp Channel for updates on global tech, business, and innovation—all in one place.

Dive deeper into the future with the Cause Effect 4.0 Podcast, where we explore the ideas, trends, and technologies driving the global AI conversation.

Got a story to share? Contact Us to reach a global audience with Impact Newswire.