Why file isolation and cleanup matter more than launch tricks for parallel AI coding agents

Running multiple AI coding agents in parallel on Windows requires careful handling of file isolation, logs, ownership, and cleanup after failed runs — not just a smart launcher setup. Each agent process needs its own set of home directory environment variables, including USERPROFILE, HOMEDRIVE, and HOMEPATH, since Windows does not natively honor the POSIX HOME convention. Shared config files like CLAUDE.md can be linked back to a global copy, but Windows symlinks require elevation or Developer Mode, making hardlinks a common fallback that can silently break 'detach' workflows. Detecting hardlinks requires comparing inode values using Node's bigint stat mode, as standard inode checks return zero on some Windows configurations and produce misleading errors. Git worktrees add another layer of complexity, since forcibly removing a worktree while a process holds an open file handle inside it will fail, requiring cleanup logic to account for locked paths.