Alexandr Wang founded Scale AI, then left to become Meta’s chief AI officer and the leader of Meta Superintelligence Labs, the group Mark Zuckerberg built after Meta’s roughly $14 billion investment in Scale earlier this year. On September 25, Wang published a personal essay on X laying out why he built Muse, an agent he describes as a “second mind” for anyone with an unformed goal. The essay does not mention Meta, Scale AI, or any company name. It is also not describing a future launch: Muse went live on September 8, and this publication reported its download numbers three days ago, which means the essay reads as a founder’s account of a product already climbing the app charts, not a preview of one still in development.
Wang’s framing centers on how little most people say out loud about what they actually want. He argues that ambitions “die by a thousand paper cuts” long before anyone tries to act on them, and that Muse exists to close the gap between wanting something and doing it. In his description, the software turns a half-formed wish into a working plan, then carries out the errands that plan requires: it drafts the outreach, places the calls, tracks down financing, and holds the person to a schedule, until, in his words, “the human part, the wanting and the dreaming, is all that’s left.” Those are Wang’s words, not a technical account. He does not explain how Muse authenticates with outside services, dials a phone number, or confirms that a task was actually finished.
That gap matters because the personal-agent category has a rough track record. Humane’s AI Pin and Rabbit’s R1 both promised general-purpose task completion in 2024, and both shipped narrow, unreliable assistants that struggled with basic requests. The hard part of this pitch has never been the vision. It is the plumbing: authentication with a hundred different services, liability when an agent books the wrong flight, and the trust required before someone lets software make a phone call in their name.
Unlike a startup announcement, Muse arrives already carrying Meta’s resources and Meta’s competitive stakes. The company has spent 2026 racing Google and OpenAI to ship agents that act across apps rather than only answer questions in a chat window. Google’s Gemini already books restaurant reservations through Astra, and OpenAI has pushed ChatGPT toward autonomous multi-step tasks. Wang now runs that effort inside Meta rather than competing against the company that bought a stake in his old one, and that changes the accountability if Muse underdelivers: it lands on Meta’s own AI leadership, not on an outside founder Meta merely backed.
The essay itself still reads like a mission statement rather than a product page. It offers no screenshot, no waitlist, and no claim about accuracy or success rate on the tasks it describes. For a product that has already shipped and already has usage numbers behind it, explaining it through prose about human potential, rather than a demo or a specs sheet, is itself a signal about what Meta thinks will sell the idea.
Operators tracking the personal-agent category should treat Muse’s early download numbers as the more reliable data point than the essay. Adoption shows whether people trust an agent enough to let it act, and Wang’s post does not. Anyone evaluating Muse against Gemini’s Astra or ChatGPT’s task agents should ask for a completion rate on multi-step tasks, since that is the number this essay conspicuously does not provide.
Reported from Alexandr Wang’s own post on X, published September 25, 2026.
