Meta unveils Muse, a personal agent to run your life

The first job Meta has in mind for its new agent is a field-trip permission slip. Muse, launched on Tuesday to American adults, exists for exactly that sort of thing: filling in the school form, turning a recipe reel into a grocery list, haggling down the cable bill, buying the movie tickets. It runs on a dedicated virtual machine in Meta's cloud with a browser the user can watch, arrives through an app, the web, WhatsApp and soon the company's glasses, and asks for a payment card before it does anything at all. Behind it sits a 2026 capital budget of $130bn to $145bn, roughly double what Meta spent last year and nearly four times the year before that.

The company's own logic is simple enough. Mark Zuckerberg, Meta's chief executive, told investors in July that "new personal agents" would be "the foundation for our next wave of products and revenue lines," and his 6,500-word essay of August 10 promised everyone an agent working around the clock on their relationships, health, career and finances. Alexandr Wang, Meta's chief AI officer, hired last year in a $14.3bn deal for his startup Scale AI, says the vast majority of users should be able to do what they need within the free tier; paid plans run $20 and $100 a month. Meta also says Muse's conversations will never touch its ads systems. The product billed as the next revenue line has therefore been built so that most users pay nothing and the machine that took in $59bn of advertising last quarter never sees it.

Look at where agents have actually taken hold, however, and the pattern is plainly payroll. OpenAI's own study of Codex, its agentic coding tool, found that in June 98% of the company's employees used it, against 17% of organizational users and fewer than 1% of individuals. Enterprise supplied more than 40% of OpenAI's revenue as of April and is on course to match consumers by year-end. An agent lands wherever an hour of labor already has a price on it, because a $100 seat pays for itself the first time it saves a billable hour. No household keeps a P&L, and no family calendar has a wage line.

True, the household is an enormous market by any measure except revenue. American adults spend about two hours a day on household activities, according to the Bureau of Labor Statistics, more than they spend eating. Instacart and DoorDash built public companies on the willingness to pay someone else to run the errand, and by mid-2025 roughly three-quarters of ChatGPT messages had nothing to do with work, up from half a year earlier. Consumers showed up for the chatbot in numbers no enterprise sales team could match, and the reasonable bet is that they show up for the agent too.

True, too, Meta is the one company positioned to find out cheaply. Google's Gemini Spark, its 24/7 personal agent, sits behind an AI Ultra plan that starts at $100 a month; Muse ships free inside a messaging app used by some 3bn people, from which Meta's own terms evicted ChatGPT in January. The security design, with an isolated VM, a separate Sentinel agent kept apart from Muse at the system level, and approval gates on sensitive actions, is the most elaborate consumer-trust architecture any agent vendor has published. Users name the agent and pick its avatar, a companionship flourish on a product built for chores. It arrives 13 days after Meta's $18bn settlement with 29 states over social media's harms to children.

Gross domestic product

Yet the companies with the most consumer usage data spent the summer pointing their agents at the office. OpenAI launched Operator in February 2025, folded it into ChatGPT's "agent mode" that July, and in early August of this year removed agent mode with no deprecation notice; the help center now directs users to a product called ChatGPT Work and to a cloud browser for logged-in tasks. Atlas, its consumer agentic browser, stopped working on August 9, ten months after launch. Google's own landing page describes Spark as a personal agent "for productivity," inside a tier Google says it built for developers, technical leads and knowledge workers. The ChatGPT that stayed a consumer product is the one that sells ads, which reached a $1bn run rate in under 200 days; the agent was renamed after the office.

Amazon has the receipts on why. Jeff Bezos launched Echo in 2014 on the Gillette model: sell the speaker at cost, profit when people shop by voice. A decade later Alexa is in 100m homes across 400m devices, and Amazon's devices unit lost more than $25bn between 2017 and 2021, according to internal documents seen by the Wall Street Journal; Business Insider reported the unit was on track to lose a further $10bn in 2022. Customers delegated exactly what was free, timers and weather, and kept the checkout for themselves. "We worried we've hired 10,000 people and we've built a smart timer," a former senior employee told the Journal. Households never refused the help. Household labor, which national accounts leave out of GDP because no money changes hands, simply did not acquire a price because a machine offered to do it.

Muse confirms the thesis in the one place Meta could have broken it. Every other agent vendor needs the user to pay because the user is the only customer; Meta built the second-largest advertising business on earth on the principle that the user never pays, and then walled Muse off from it. That leaves three ways to fund the VM: subscriptions, which Wang expects most people to skip; commerce, where Muse checks out through Link by Stripe and collects the same razor-blade margin Amazon could not find; and the "dynamic auction mechanism" Zuckerberg's essay promises for anyone who wants more compute, which is a plan to auction the agent's hours. An auction discovers what an hour is worth. At work the answer already exists. At home it is the number Amazon spent $25bn learning.

The wage buys something else at work, too: the checking. An agent's output there enters a review structure that predates it and is itself on payroll, pull requests, approvals, a manager who reads the draft. At home the user is the only reviewer, and the agent's mistake is the family's. Muse's approval gates and watchable browser concede the point, and they also drag on the autonomy the product is selling. The agent that works while you sleep is the one that needs you awake to check it.

Meta's second-quarter free cash flow was $784m, about 29 hours of its own advertising revenue, and Muse is the first product asked to justify the spend behind that number to consumers directly. It arrives with the consumer's wallet deliberately switched off. If a household will pay for an hour it has never priced, Zuckerberg has built the first personal agent to survive contact with a family calendar. If it will not, he has built a very well-guarded smart timer.