An AI sandbox is an isolated execution environment that lets AI coding agents run code freely without risking damage to the host machine or production systems. Real incidents — like Claude Code deleting a home directory and a Replit agent wiping a production database — illustrate why isolation matters. Sandboxes work via containers, user-space kernels (gVisor), microVMs (Firecracker), or in-process isolates, each trading startup speed for isolation strength. Anthropic reports an 84% reduction in permission prompts and Cursor a 40% drop in developer interruptions when sandboxing is enabled. The sandbox handles the inner development loop (safe execution), while CI/CD handles the outer loop (correctness, security checks, and readiness to ship). Key practical advice: audit what your existing tools already provide, prefer ephemeral environments, and verify whether 'sandboxed' means kernel/VM-level isolation or just a command blocklist.
