Meta introduces Muse, a personal AI agent

Meta introduced Muse on September 8, 2026, describing it as a personal agent for helping people pursue goals and suggesting ideas.[1] The company says the agent can turn a goal into a personalized plan and advance the work on its own, rather than stopping at an answer.[1] Muse can open a browser, fill out forms and negotiate on a user’s behalf.[1] Meta says it keeps working after a user closes the app, then returns when circumstances change or it needs approval.[1] People can use Muse in its app or directly through WhatsApp.[1]

Muse asks for approval before taking actions such as sending an email or making a purchase.[1] Meta says internet access also requires Sentinel approval, requested when needed, and users choose which apps Muse connects to and the amount of access it receives.[1] For checkout, Muse can use Link, Stripe’s payment product.[1] Meta says Muse is the first AI agent covered by Link’s purchase protections; Link’s agent wallet creates a one-time-use card to keep a user’s real card details hidden.[1] These controls make approval points and payment handling important considerations for teams assessing delegated tasks.

Meta said Muse was rolling out in the United States on iOS, Android and the web.[1] It said the agent would be free for most needs, with subscriptions for people who want to do more.[1] Meta says Muse runs on a dedicated virtual machine that houses both the agent and a person’s data.[1] The company says Muse does not share users’ conversations or virtual-machine data with its advertising systems, and users can opt out of having their interactions used to train its AI models.[1] Meta said it would introduce Muse Confidential VM later in 2026, encrypting the virtual machine with a key held only by the user.[1]

Tokenpost reported that Nat Friedman’s OpenClaw experiments in February 2026 helped convince Meta’s team that personal agents could reach consumers beyond coding workflows.[2] Wang said Meta built an initial Muse prototype in one to two weeks and demonstrated it to the company’s board.[2] He said the product launched seven months later, after a lengthy effort to make it reliable.[3] The team evaluated more than 100 capabilities and behaviors.[2] Meta reported that earlier model versions passed about 80 checks and failed 20, while a specially trained Muse Spark 1.3 checkpoint passed the full evaluation.[2] Wang said fewer than 200 people worked across model and product development.[2]