Claude Code's cloud sandbox boots a custom kernel in 430ms

Rohan Adwankar took apart the virtual machines behind two hosted coding-agent platforms and wrote up a teardown of both platforms. Claude Code runs on Firecracker KVM guests booting kernel 6.18.5-fc-v20, with a Rust and Tokio binary called process_api standing in for PID 1 instead of systemd. That binary reaches init in about 430 milliseconds, and the full harness is ready in 6.4 seconds. The guest gets a 256GB writable disk, separate read-only disks for the harness and tools, OAuth tokens the host mints fresh on every boot, and a PID 1 flagged non-dumpable so root inside the VM can't read its own memory.

Instinct takes a different bet. It rents E2B sandboxes running stock Ubuntu 22.04.5, cold-booting a full XFCE desktop in about 1.26 seconds. Agent memory lives as a git repo of Markdown files linked with wiki-links, searched with grep, and pushed to S3 as bundles authenticated with short-lived STS credentials.

The two setups answer the same question with opposite answers: how much should you trust code a language model wrote, enough to let it run unsupervised? Claude Code assumes a hostile guest and builds a custom kernel and init system to contain it. Instinct assumes a disposable VM, leans on the isolation E2B already provides, and spends the savings on a real desktop environment and a memory store a human can read.

Anyone renting either platform has a concrete boot budget to measure against: 430 milliseconds to a minimal init versus 1.26 seconds to a full desktop, plus two different answers for how much a coding agent should be allowed to see about the machine it's running on.