Alexandr Wang tells David Senra how Muse went from a 2025 memo to a product in seven months

Meta chief AI officer Alexandr Wang has given the most detailed public account yet of how Muse was built, saying the personal AI agent went from a working prototype to launch in about seven months, with the prototype itself taking “one to two weeks” before it was demoed to Meta’s board. In a new interview with podcaster David Senra, Wang traced Muse back to a June 2025 “Personal Superintelligence” memo written with Mark Zuckerberg, Nat Friedman, and a few others, and credited OpenClaw experiments inside Meta in February with convincing the team that consumer personal agents were ready.

“It is not a chatbot. It is an agent,” Wang said, restating the framing that has defined Muse since its Sept. 8 launch. Fewer than 200 people worked across the model’s and the product’s development, he said, and when pressed on user numbers he would only repeat the company’s line: “We have said millions.”

The memo and the fork in February

According to Wang, Meta Superintelligence Labs began with the June 2025 memo, which he said sketched out what Muse could become before AI agents had meaningfully arrived. The bet took shape this February, when two things converged: OpenAI’s Opus 4.5 showed what agents could do, and Friedman — then head of Meta’s AI-products division — went all in on OpenClaw inside the company.

Wang said he installed OpenClaw himself and tried Friedman’s deep psychoanalysis prompt, feeding it his email and photo access. “It was maybe the equivalent of three years of therapy compressed into this one moment,” he said. Zuckerberg, according to Wang, was running an agent at home with his kids and using one to review his own MMA training video for tips, while February staff meetings turned into show-and-tells of agent experiments.

That month also settled Meta’s strategy. The conventional wisdom then, Wang said, was that Anthropic and Claude were far ahead on coding agents and that coding was the only worthwhile form factor. Meta bet the opposite way: personal agents for everyday tasks, because Meta already had the consumer distribution to make them work. Meta’s apps touch about 3.5 billion people a day, Wang said, and those products are built around what people want to do rather than what they have to do.

100 checks, all green, before launch

The rest of the seven months, Wang said, was spent turning a fragile demo into a reliable consumer product. A prototype was assembled in a week or two, shown to the board (with the Jolly mascot bot and an early version of Muse Charm already in the mix), and then hardened until September. The team logged more than 100 specific behaviors the model needed as a personal agent, built an evaluation for each, and ran review cycles until the failures cleared. Earlier model checkpoints passed about 80 and failed 20, Wang said; the first checkpoint that came back green across the board was a specially tuned version of Muse Spark 1.3.

Wang was blunt about what decides Muse’s future now: reliability at scale. OpenClaw’s own trajectory, he argued, showed what happens when reliability never arrives. “We had our first model checkpoint which was green across the board,” he said, framing the launch gate as the hard part, not the prototype.

The interview also surfaced one roadmap item: Wang said Meta plans to introduce a Muse Confidential VM later in 2026, with encryption keys controlled by users.

Why it matters

Muse users get the most detailed look yet at how Meta decides what the agent is allowed to do. The 100-plus behavior evaluations are the company’s definition of a trustworthy personal agent, and the reason Meta said no to an earlier launch. It also spells out the product’s organizing bet: the winning personal agent will be judged on reliability at everyday tasks, not on benchmark scores, and distribution through apps people already use matters more than being first to coding.

Builders and competitors will pay closest attention to the OpenClaw thread. An open-source agent project, first tried as a February experiment, set the direction for a product Wang describes as central to Meta’s AI strategy. Friedman’s move (buying hundreds of Mac minis for the team, by his own account on X earlier this year) looks less like a curiosity and more like the point at which Meta committed.

The full interview, “Creating Muse: The Fastest-Growing AI Product Since ChatGPT,” is on YouTube and at davidsenra.com/episode/alexandr-wang. Wang shared it on X as a 72-minute conversation covering the Muse story, meme marketing, and his own path from Scale AI to Meta.