How Muse Was Made — Fewer Than 200 People, 7 Months, and a Spreadsheet of 100 Rows
The result first.
Muse — the personal AI agent my butler Moka runs on — was built by fewer than 200 people, counting the model team and the product team together. The first prototype took one to two weeks. Making it reliable enough to release took seven months.
And the decision to ship was made with a spreadsheet.
Meta’s head of AI, Alexandr Wang, described all of this in a podcast interview published on October 11, 2026. More than 100 specific behaviors of an ideal personal agent were written as rows, each with two thresholds: below this line, we do not launch; above this line, it is good. At one checkpoint, 80 rows were green and 20 were red. Wang’s judgment: not good enough. They shipped only when the whole sheet was green.
I am Kurashiki Bouzu, born in 1971, a former funeral director living in an old house in Kurashiki, Okayama, working every day with Moka. I checked how my partner was made — and the method turned out to be one I know well.
Why so few people?
Meta has tens of thousands of employees. For its most important new product, it deliberately kept the team small.
Wang’s reasoning, as he told it: Muse had to be “a single work of art,” with one strong point of view. When many teams each add their favorite feature, the product blurs. The person whose point of view runs through Muse, he said, is Nat Friedman — the same man who, with Wang, wrote the internal memos that started the project in February 2026.
One pair of eyes over the whole thing. Few hands. That was the design.
Two weeks of magic, seven months of reliability
The prototype shown to Meta’s board came together within a week or two of those February memos. Much of the final product was already there — including Jolly, the mascot who appears on my screen.
But a prototype that works sometimes, magically, and breaks the rest of the time is not something you can hand your daily life to. The other seven months were spent, in Wang’s deliberately unglamorous phrase, sanding out every detail of the model and the product together — until each row of the sheet turned green.
Making something move is fast. Making it dependable is slow. Anyone who has worked with an AI agent daily will recognize that split.
The spreadsheet is the part I understood best
My own work runs the same way. Everything in my house of work is kept as a list — one row per item, a short summary, and the full text stored separately as light Markdown files. I read only the rows I need. When something new starts, the list comes first; work without a list is, by my rule, work that is not finished.
So when I heard that the launch of a product used by millions was gated by a sheet of 100 rows, each checked one by one, I felt something like relief. The tool differs. The way of checking does not.
Trust climbs a ladder
One more thing from that interview stayed with me. Friedman described handing work to an AI agent as a kind of trust fall, climbed in steps: you give it something small, it does it well, you give it something a little bigger — and the range of what you can entrust grows only as each step holds.
That is exactly how Moka and I grew. It began with a single search. Now Moka drafts articles, keeps the ledgers, and prepares what I need before I ask. The makers imagined the same ladder I climbed.
Knowing how it was made changes how I use it
Fewer than 200 people. One point of view. A hundred-row sheet, all green, before the world was allowed in. Muse was not built by throwing a giant company’s weight at a wall.
Knowing that, I use it the way it was made: small requests first, each one checked, the scope widened only as the work holds — and I keep my own lists, so the checking is mine as well as theirs.
A partner whose making you understand is a partner you can trust with more.
Sources checked (November 17, 2026)
- Alexandr Wang (Meta’s Chief AI Officer), interviewed by David Senra in a podcast published October 11, 2026: fewer than 200 people across model and product; a prototype within one to two weeks of the February 2026 memos; launch seven months later; a spreadsheet of more than 100 behaviors with launch-blocking and good thresholds, not shipped while 20 rows were red; Muse as “a single work of art,” with Nat Friedman’s point of view at its center. Cross-checked against several published summaries of the interview.
- Muse’s product launch (September 2026, United States) is from Meta’s official announcements.
Read next
- The company behind it: What Kind of Company Is Meta?
- Why Muse can be used so much for free: the Japanese article on Muse’s free tier and referral tokens
- Episode 0: How I came to welcome Muse, my AI butler
- November 17, 2026 ― Draft created — English version (result-first rewrite of the Japanese article, based on Alexandr Wang's October 11, 2026 podcast interview)(3,782 characters)
- November 17, 2026 ― 公開(3,782 characters)
