Creating Muse: Alexandr Wang on the Seven-Month Sprint to a Personal Agent

One internal memo sets the compass for AI work at Meta. In June 2025, Mark Zuckerberg announced the new Meta Superintelligence Labs unit with a public note: personal superintelligence for everyone. The thesis was plain; frontier models should act as leverage that improves each person's life rather than as a central mind that takes people's work away. CNBC published the memo, which already signaled that the lab would aim at consumers instead of joining the coding race. I read that sentence as the first stone on the road to the Muse launch a year later.

The road to that compass passed through a disappointment. Llama 4 had failed to deliver the jump the company expected, and management felt it was falling behind in the AI race. The answer was drastic: a 14.3 billion dollar partnership around Scale AI, the appointment of its founder Alexandr Wang as Meta's first chief AI officer, and former GitHub chief Nat Friedman taking charge of products. Fortune described the reorganization that followed, paired with a compute budget running into the hundreds of billions. The Wall Street Journal reported packages reaching into the hundreds of millions to pull names from rival labs. Meta had decided to rebuild an AI program that looked damaged into something new.

The mood across the industry in early 2026 was different. Anthropic's Opus family had opened a clear lead in coding, and nearly every lab treated the coding agent as the one capability that mattered. On Wang's account, the belief inside Meta was that joining that crowd would be a mistake. Coding, in his framing, is the peak of work-oriented AI; Meta's heritage is what people want to do. That distinction became the core of the strategic bet that turned the lab's model roadmap toward personal agents.

February 2026: a prototype between fear and euphoria

The break came from an unexpected corner. OpenClaw, a public agent project by Austrian developer Peter Steinberger, went viral in the first weeks of 2026, and a long Lex Fridman conversation published on YouTube carried it into the mainstream. Friedman was among the first to try it and described the feeling as somewhere between terrifying and euphoric: he had told the agent he wanted to drink more water, and it watched him over home security footage to check whether he did, praising him when he complied. As The Decoder later wrote, moments of that kind showed what the personal-agent idea promised even in raw form.

Wang crossed a similar threshold in the same weeks. He ran a deep self-examination prompt suggested by Friedman on his own agent; the agent looked at his mail, his photos and records it could find, asking layered questions round after round. Wang compares that single session to three years of therapy. What shook him was how the agent fused scattered personal data into accurate conclusions. Both men reached the same verdict in those days: the sparkle inside a rough experiment could redirect consumer technology.

Before February ended, two separate notes went up to the board. Wang argued the super agent was a destination for consumer products; Friedman wrote about trust as a precondition and the principles agents must obey. Meanwhile Zuckerberg himself was trying the product at home: chores with the kids, household organization, and an agent watching his martial arts training footage to suggest technique fixes. On Wang's account, staff meetings of that period had a fixed ritual where everyone recounted the wildest thing they had done with their agent. Having top management in the user seat gave the project institutional cover.

Once the decision landed, the team assembled a working prototype in under two weeks and demoed it to the board. The playful helper character and the first version of the keychain-sized charm shown that day formed the core of today's Muse; not every tab existed, yet most choices that define the final product were already in that first demo. The point Wang stresses is what came between that demo and launch: seven months spent turning magic into reliability. An experiment that amazed at times and broke most of the time had to become a product trustworthy enough for billions.

Seven months of sanding: model roadmap and 3.5 billion reach

The team locked model work onto personal-agent behavior. More than a hundred desired behaviors of the ideal agent were listed one by one; a measurement setup was built for each, and weak spots were cleared in periodic reviews. Wang calls this sanding: Meta's habit of measuring and improving every step of conversion and onboarding flows was now applied to the model itself. He believes this is also why OpenClaw flared and faded; the magic moments existed but the experience was unpolished, and Meta had the patience and the compute to do the polishing.

Why Meta, then? Wang answers on two levels. The first is heritage: the company grew by connecting people with their loved ones and interests, which is the realm of things people want to do, not things they must do. The second is reach: 3.5 billion people open the company's apps every day. His reading of the general AI promise is that hours spent on obligatory work fall toward zero while desired work expands; Meta's products already occupy that second realm. While coding and office productivity belong to others, personal consumer agents were Meta's home ground. Meta's own statement makes the same point: power should sit in people's hands so they can direct it at what they value.

The first fruit of that thesis on the model side arrived in April 2026. Muse Spark was presented as the lab's first reasoning model: a fast mode for everyday tasks and deeper thinking modes for hard math and science questions. The measurements reported by Fortune showed a competitive but not leading picture; on a doctoral-level reasoning test it scored 89.5 percent against 94.3 for Gemini 3.1 Pro, and the company openly admitted gaps in coding and long-horizon agent work. Axios wrote that the effort under the Avocado code name had taken nine months and that an openly licensed version was planned. For Wang what mattered was not the crown but the first rung of the ladder toward personal agents.

The product curtain rose on September 8. Codenamed Hatch inside the company, Muse was introduced as an agent taking over digital chores like booking appointments, filling forms and watching home cameras; CNBC reported a feed assembled from Facebook and Instagram connections, an ideas tool for trips, and a lean version running over WhatsApp. Users could name their agent and pick its look. The app shipped in the US on iOS, Android and the web, with Canada following days later; glasses support was left for later. Wang says the front was kept deliberately simple while advanced coding work ran behind the scenes.

The numbers showed the social giant could win a download race. Sensor Tower data relayed by CNBC counted 730,000 downloads in the first five days, Reuters pointed to 55 percent average daily growth across the first ten, and September 19 set a daily record at 264,000. The Apptopia comparison published by TechCrunch was more striking: 2.8 million installs in the first 12 days, 1.8 million against 1.3 million for ChatGPT on iOS alone in the US and Canada, and a daily active lead of 642,000 to 231,000. Sensor Tower data relayed by Yahoo logged the 5 million mark in 22 days; ChatGPT had needed 56, Grok 103 and Claude 492. The app sat at the top of both app stores.

Price, privacy and the copying debate

Meta opened with a free tier and two paid tiers at 20 and 100 dollars a month for heavy users. Wang told CNBC most people would stay on the free tier while the upper tiers covered compute costs. There is no advertising in the app; the company is sounding out a cut of shopping done through the agent. As Yahoo noted, this is an attempt to convert a download triumph into subscription and transaction revenue. Behind the agent's friendly face sits a classic platform question: how does attention turn from impressions into a cut?

The thorniest chapter of the launch was privacy. Wang stressed the agent runs in an isolated environment inside company infrastructure, never sees passwords or payment details, and asks permission before anything sensitive. Product vice president David Singleton explained users can switch off the use of agent interactions for model training; for those who do not, critical identifying information is stripped out. CNBC recalled the dark backdrop: the company had paid a 17 billion dollar settlement with states over claims it understated harm on its apps, while agent security risks and data-center backlash filled the agenda. Trust is not a feature of this product but a precondition.

One admission made the launch more interesting. When early users noticed Muse file names and behavior texts overlapping OpenClaw almost word for word, Friedman wrote without hesitation: heavily inspired as a product, yet built from scratch. The Decoder reported the SOUL personality file was nearly identical; asked about it, Friedman answered that Steinberger had gotten those things exactly right. In one founder's viral phrase, Muse was OpenClaw for ordinary users. TechCrunch likened the move to a familiar play: as with Stories borrowed from Snapchat, carry the best parts of a promising product inside.

Competition is not standing still. Yahoo reported OpenAI introduced Dots, a platform entering through the enterprise door with a cute avatar, with a consumer version on the horizon. But for Wang the real test is not downloads but retention: distribution wins the race, only product wins the habit. The data point highlighted by TechCrunch gives pause; 95 percent of users are Facebook users and 63 percent use Instagram. Yahoo, citing Sensor Tower, wrote the company devoted half its in-house promotion impressions to Muse in mid-September. The wind is at its back, yet the product itself must fill the sails.

The most striking frame for product craft comes from the recording date itself: the conversation was taped on the 15th anniversary of Steve Jobs's death. Wang's admiration for Jobs connects to the single-point-of-view product; as Apple is its founder's taste living on in ten thousand people, Muse should speak in one voice for the lab's idea of personal intelligence. That is why the team obsessed over delivering the wow moment as early as possible; the charming accessory, the playful helper and the tabs all serve that thought. The crowd distribution brings must be greeted at the door by a refined first minute.

The philosophical vein of the conversation is agency. Wang argues modern life wears people's ambitions down; the promise of AI is an escalator of agency that starts with childhood dreams and keeps setting bigger ones with each success. Otherwise, he warns, we turn into crowds that surrendered initiative. The schedule ahead holds glasses support, a public model release and shopping-cut experiments; on the horizon stands the sentence from the 2025 note: personal superintelligence. I find the real value of this recording is not a launch but a lab's change of character on the record; download records come and go, that change may last.