Isolate multiple AI agents in a single Vercel Sandbox by giving each its own Linux user and sharing files through a group directory.

Documented steps

1. Create a sandbox as usual: `const sandbox = await Sandbox.create()`. Docs: https://vercel.com/docs/sandbox/multi-agent
2. Create a Linux user per agent: `const alice = await sandbox.createUser('alice')`. Username must match `/^[a-z_][a-z0-9_-]*$/` and be at most 32 characters; the user gets `/bin/bash` and a private home directory at `/home/alice`.
3. Run commands or file operations scoped to that user via the returned `SandboxUser` handle (methods mirror the sandbox's own `runCommand`/file APIs but execute as that Linux user).
4. To reference a user that already exists (e.g. restored from a snapshot, or `root`) without creating it, use `sandbox.asUser('bob')`.
5. For shared workspace between agents, create a group: `const devs = await sandbox.createGroup('devs')` (shared directory at `/shared/devs`, setgid so files inherit the group). Add members with `await sandbox.addUserToGroup('alice', 'devs')`; revoke with `sandbox.removeUserFromGroup('alice', 'devs')`.
6. Group revocation applies immediately to new commands, but processes already running as that user retain group access until they exit.