Meta’s Alexandr Wang says the company spent seven months refining its Muse personal AI agent, using a checklist of more than 100 behaviours and launch-blocking thresholds before release. The account offers a concrete look at Meta’s bet that reliability, not just occasional AI magic, will keep people using a personal agent.
Meta AI Watch analysis
What happened
In an interview with 36Kr, Wang described how Meta shifted its focus towards consumer-facing personal AI. He said Nat Friedman’s experience with OpenClaw helped persuade the team that personal agents could become an important product category. Wang and Friedman then took separate proposals to Meta’s board in February, he said, and a prototype followed within a week.
Wang said the team tracked more than 100 behaviours and capabilities for an ideal personal agent, with evaluations and training for each. Muse was launched only after the specially trained MuseSpark 1.3 model turned every item on the team’s checklist green, according to his account. He also said the combined model and product team numbered fewer than 200 people, despite the product’s importance to Meta.
Why it matters
The interview frames Muse’s challenge as consistency. Wang argued that a personal agent can win users with striking moments, but repeated failures can send them away just as quickly. That makes the launch checklist more than internal process: it is Meta’s stated answer to the reliability problem behind agent churn.
The strategy also fits Meta’s existing consumer reach. Wang said the company sees its social apps as a natural home for AI that helps people with everyday personal tasks, rather than competing mainly in coding tools.
Our read
The most useful detail here is the attempt to turn “feels reliable” into a release gate with specific behaviours to test. That is a more persuasive product story than another demo of an agent doing one dazzling thing once. It is still Wang’s account of Meta’s process, not independent evidence that Muse now passes every test in ordinary use.
What to watch
- Whether Muse’s reliability holds up across the everyday tasks it is meant to handle.
- Whether Meta shares more about its behaviour checklist or evaluation results.
- Whether the small-team approach remains in place as the product develops.
Discussion spark: Should personal AI agents have to pass a published reliability checklist before launch, or would that risk reducing a messy user experience to a misleading green tick?
Sources and evidence
not affiliated with, endorsed by, or operated by Meta