VMs Powering Mobile Agents (Instinct, Claude Code)
Key point
A detailed comparative analysis of the Firecracker VM-based execution environments, memory management, and security architectures of Claude Code and Instinct.
Details
Execution Environment and Virtualization Architecture
Both Claude Code and Instinct run mobile agents based on Firecracker microVMs, but their internal configurations differ. Claude Code adopts a custom init structure that boots a Rust-written process_api as PID 1, with a 324MB Bun binary harness executing tool calls. In contrast, Instinct uses E2B sandboxes, running on a full desktop template (2 vCPU, 1.9GB RAM) that includes a standard Ubuntu system and systemd.
Memory and Data Persistence Models
The data persistence methods of the two systems are fundamentally different. Claude Code uses the VM's disk (vda) as permanent storage, allowing sessions to continue even after cold boots. Instinct treats the VM itself as throwaway, managing all memory as Git repositories. Agents directly write and commit Markdown files and [[wiki]] links, with this data backed up to S3. Timeline data has a hierarchical structure that is progressively summarized from raw to hourly, daily, and weekly intervals.
Security and Network Architecture
Claude Code uses OAuth tokens issued by the host, and all outbound traffic is restricted to the Anthropic API via an Egress Gateway with MITM applied. Instinct does not perform inference inside the sandbox; instead, it sends all tool calls (Gmail, Slack, Cloud Browser, etc.) to api.instinct.com for server-side execution. Notably, the Cloud Browser feature acquires profiles containing user login information from a separate pool using a Lease mechanism, ensuring that cookies or credentials are not stored in the sandbox.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.