Meta says Muse took seven months to ship, with fewer than 200 core staff and a 100-plus test gate

Meta Chief AI Officer Alexandr Wang has outlined how the company built Muse, a personal AI agent designed to handle tasks for users. According to Wang, the product moved from prototype to launch in seven months, with a core model and product team of fewer than 200 people. He said Meta chose to focus on personal agents even as Anthropic was seen as far ahead in coding agents earlier this year. Wang and Meta AI product head Nat Friedman believed Meta’s advantage lay in building AI products for mainstream users across its apps, which Wang said reach about 3.5 billion people daily. After both executives tested OpenClaw in depth in February, they submitted memos to Meta’s board. One memo from Friedman, titled Trust Is a Must, argued that personal agents had to earn user trust. The team then produced a Muse prototype within one to two weeks, but the early product was unstable. Over the following months, Meta defined more than 100 required capabilities and built tests for each one. Some model versions passed about 80 tests, but failed roughly 20, delaying release until the specially trained Muse Spark 1.3 cleared the full evaluation set.