Inside the VMs Where Mobile Agents Run

Rohan Adwankar published a breakdown of the virtual machines where Claude Code (Anthropic) and Instinct mobile agents run: both products use Firecracker microVMs, but with opposite architectures — a sealed “operator” inside the guest for Claude Code and a pure execution sandbox without a model for Instinct.

Compiled by Sergey KostenchukPublished 2026-09-08Updated 2026-09-18

2026-09-08 Coding Anthropic

💻 Inside the VMs of Mobile Agents

Rohan Adwankar broke down the machines where Claude Code (Anthropic) and Instinct run. Claude Code: Firecracker microVM, custom kernel 6.18.5-fc-v20, PID 1 — a sealed Rust/Tokio process process_api (vsock 2024), vda disk 256 GB, inference SSE via a MITM gateway. Instinct: E2B sandbox (2 vCPU, XFCE) on the same Firecracker; no model in the box — tool calls go out via GraphQL, memory is a Markdown git repository that the agent pushes to S3.

🌍 One of the first public reverse-engineering reports on agent isolation: one Firecracker — two architectures: an “operator” inside the guest versus an execution surface with memory separate from the machine.

👤 Kernels, ports, disks, and GraphQL instead of abstractions; the ws-term repository is open — you can enter the VM and compare approaches.

Source 1: https://rohanadwankar.github.io/posts/platforms.html Source 2: https://news.ycombinator.com/item?id=49605644

Sources