A live stream walkthrough comparing different sandboxing approaches for AI coding agent harnesses (Claude Code, Codex, Cursor, Copilot). Covers four isolation tiers: built-in harness sandboxing (Claude Code's /sandbox using macOS Seatbelt or Linux Landlock/Bubblewrap), third-party host-native tools like nono.sh (from the Sigstore team), traditional Docker containers, and Docker Sandboxes (SBX CLI) which spins up per-project micro-VMs. Key motivations include enabling 'yolo mode' (dangerously-skip-permissions) safely, controlling which CLI tools and API keys agents can access, and avoiding agents accidentally touching infrastructure credentials. Trade-offs discussed: built-in sandboxing is easiest but least isolated and model-breakable; nono.sh adds kernel-level file system restrictions without containers; Docker containers offer strong isolation but complicate Docker-in-Docker scenarios; Docker Sandboxes provide the most complete isolation at the cost of heavyweight VM overhead.