Singleton describes the cloud computer as a "full-fledged" machine where users can install software, write and compile code, or browse the web, aiming to replicate the experience of a physical machine under a desk. The service includes a classic file explorer view and a "Download your agent data" option for exporting user content, reflecting a deliberate choice for technical transparency. Singleton notes that while Meta was an early player in the AI race with its Llama models, it has not kept up with frontier models from OpenAI or Anthropic, suggesting that Muse's success is based on product reach rather than model performance.

The architecture separates the user workspace from sensitive system components using a "Sentinel" process that monitors sensitive actions, with passwords and credentials stored outside the "Runtime Cell". This design ensures that even if a user's environment is compromised, it cannot access privileged system components or other users' data. Singleton emphasised that the Runtime Cell has its own root filesystem, providing an additional layer of security against prompt injection and unauthorised access to Meta's infrastructure.

Muse has seen rapid adoption, pulling in over 500,000 users in its first week and reaching number one in the Apple App Store. Singleton stated that Meta is betting that the best product with the widest reach matters more than the best model, a strategy that appears to be paying off despite the company's lagging performance in frontier model development.

Meta has faced some friction with external services, as Amazon blocked Muse from accessing its online store due to privacy and security concerns over the agent failing to identify itself and seemingly capturing customer credentials. This incident highlights the ongoing challenges AI agents face in navigating third-party platforms and the need for clear identification and permission protocols.