Meta Focuses On User Control For New AI Assistant Muse

Meta wants users to feel they are always in control of its new AI assistant, Muse, according to Vishal Shah, the company's vice president of AI products. Shah discussed the assistant during Meta Connect 2026, where the tech giant also showed off new smartglasses and virtual reality glasses. Meta launched Muse several weeks ago and has rolled out updates almost daily since then. Recent additions include a Mac application that lets Muse control a computer directly, along with voice and embodiment features for real-time conversation. Meta also plans to make Muse the default AI agent across its line of smart glasses.

Autonomy and Permissions

Shah explained that how much freedom an AI agent should have is a key challenge for the industry. Certain actions, such as making a purchase or sending messages, always require explicit user approval. For other tasks, permissions adjust as Muse learns a user's preferences over time. The system starts off conservative and can either ask each time or act alone later depending on the situation. For example, when connected to Gmail, Muse defaults to read-only access until the user grants permission to send or write emails. Shah noted that this cautious approach avoids unwanted surprises while preventing the assistant from asking too many annoying questions.

Technology and Future Plans

Muse runs on a new multimodal reasoning model called Muse Spark 1.3. This model is built for personal superintelligence tasks, such as managing long-running goals and coordinating multiple subagents, rather than general benchmark performance. Its ability to write code lets it build connectors for apps that lack built-in integration. The assistant is currently available only in the United States and Canada. Shah stated he is very excited about the Indian market due to its strong demand from small businesses and individuals, though expansion will require local app integrations and language support.

Context

AI assistants are increasingly moving from basic text-box chatbots to proactive agents capable of executing tasks across apps and operating systems. Managing privacy, security, and user permission limits remains a central design hurdle for tech companies building these systems. Meta integrates these controls directly into its cloud virtual machine infrastructure to ensure safety as agents take on more autonomous duties.