Meta’s Muse Grew From OpenClaw Experiments Into a Personal AI Agent

Meta’s Muse personal AI agent grew from OpenClaw experiments into a consumer product designed to complete everyday digital tasks, after about seven months of reliability testing and product development before its U.S. launch.

Meta Chief AI Officer Alexandr Wang said the company built an initial Muse prototype in one to two weeks and demonstrated it to Meta’s board. Muse launched in the United States on Sept. 8 for iOS, Android and the web.

“It is not a chatbot. It is an agent,” Wang said.

Muse can send emails, book travel and fill out online forms. It can continue working after the app is closed and pause for user approval when required, including before sensitive actions such as sending an email or making a purchase.

The system operates inside a dedicated cloud computer called Muse Secure VM. A separate Sentinel system must approve internet access, while users control which applications Muse can access. Meta also provides audit trails of the agent’s activity and separates Muse data from its advertising systems.

The product grew from Meta’s broader Personal Superintelligence vision, which Wang described in a June 2025 memo. Nat Friedman’s OpenClaw experiments in February 2026 helped convince the team that personal AI agents could reach consumers beyond coding workflows.

“It was maybe the equivalent of three years of therapy compressed into this one moment,” Wang said, describing his experience with OpenClaw.

Reliability testing became a major part of development. The team evaluated more than 100 capabilities and behaviors. Earlier model versions passed about 80 checks but failed 20, while a specially trained Muse Spark 1.3 checkpoint passed the full evaluation.

“We had our first model checkpoint which was green across the board,” Wang said.

Fewer than 200 people worked across Muse’s model and product development, Wang said. Meta describes Muse Spark as its most capable model for real-world agentic work.

Meta’s consumer focus reflects the reach of its existing products, which have about 3.5 billion daily users across its apps. The company has disclosed only that Muse has “millions” of users.

“We have said millions,” Wang said when asked for a more precise figure.

Meta plans to introduce Muse Confidential VM later in 2026, with encryption keys controlled by users. Muse is also planned for Meta’s AI glasses, extending Meta’s launch of Muse as a personal AI agent toward additional hardware.

The product’s documented reliability testing, permission controls and approval requirements show Meta’s effort to move AI beyond conversational answers toward completing digital tasks for users.