Peter Zhang
Jan 14, 2026 16:05
Manus details its cloud sandbox system powering AI agent execution, featuring Zero Trust security, persistent file storage, and 24/7 autonomous operation.
Manus, the AI agent startup acquired by Meta in late 2025, has published technical documentation revealing how its sandbox infrastructure enables autonomous task execution—a system that now falls under Meta’s expanding AI portfolio.
The Virtual Machine Approach
Unlike chatbots that simply generate text, Manus allocates a fully isolated cloud virtual machine for every user task. Each sandbox runs independently with networking, file systems, browsers, and development tools. The AI can write and execute code, build websites, and create mobile applications—all running on remote infrastructure rather than consuming local resources.
“The power of Sandbox lies in its completeness—just like the personal computer you use, it has full capabilities,” the company stated in its January 14 documentation release.
Lifecycle and Retention Limits
The sandbox operates on a sleep/wake cycle. When inactive, it hibernates automatically while preserving files. Users returning to tasks trigger automatic restoration. However, dormant sandboxes eventually get recycled—after 7 days for free users, 21 days for Manus Pro subscribers.
Upon recycling, the system creates a fresh sandbox and restores key files: user uploads, AI-generated artifacts, and project files like presentations or web applications. Temporary code and intermediate files don’t survive the transition.
Zero Trust Security Model
Manus employs Zero Trust architecture, granting users and the AI agent full control within their sandbox—including root access and system file modification. The trade-off: any destructive action stays contained. Sandbox operations can’t access user account data or impact other sessions. If something breaks catastrophically, Manus spins up a replacement automatically.
The company drew a sharp distinction between “sharing” and “collaboration” modes. Sharing a task only exposes conversation logs and output artifacts—the sandbox remains invisible. Collaboration, however, grants partners full sandbox access. They can instruct the AI to read or modify any files present, creating potential data exposure. Connectors to external services automatically disable when collaboration begins.
Meta’s Broader AI Ambitions
Meta’s acquisition of Manus, finalized in late 2025, brought this autonomous agent technology in-house. The company clarified concerns about Manus’s origins, with reports noting Meta addressed questions about the startup’s “Chinese connection” at the time of purchase. Manus was originally developed by Monica before the Meta deal closed.
The sandbox feature remains available across all subscription tiers, suggesting Meta intends to maintain Manus’s existing user base while integrating the underlying technology into its broader AI infrastructure. For developers and power users, the persistent cloud execution model offers something most consumer AI tools lack: the ability to run complex, multi-step tasks without keeping a browser tab open.
Image source: Shutterstock
Source: https://blockchain.news/news/meta-manus-ai-sandbox-architecture-autonomous-agents
