The VMs Powering Mobile Agents (Instinct, Claude Code) (rohanadwankar.github.io)

Recent advancements in AI mobile agents have led to significant changes in how platforms like Claude Code and Instinct operate, transitioning from local computer environments to virtual machines (VMs) accessible on mobile devices. Claude Code utilizes Firecracker microVMs with custom operations written in Rust, ensuring secure and efficient execution. Its architecture allows agents to interact seamlessly with users by managing persistent and shared storage separately, enhancing the continuity of user experience even when compute resources are reclaimed. The system emphasizes strong security measures, such as isolating its control process and using secure communication protocols. On the other hand, the newly launched Instinct leverages a unique approach by utilizing rented E2B sandboxes—disposable Ubuntu environments optimized for rapid deployment. Instinct focuses on memory management, storing user interactions in a Markdown-based git repository that reflects human memory dynamics through structured timelines. Notably, it processes tasks by executing commands in a secure environment while delegating inference and tool interactions to remote servers through GraphQL. This design allows Instinct to maintain user context efficiently without burdening its local sandbox, marking a shift toward more sophisticated and integrated AI assistants.