A hands-on evaluation of Windows 11's Copilot Actions and Agent Workspace, written by someone who actively uses agentic AI workflows in their home lab. The author enabled the feature, tested it, then disabled it due to concerns about the default trust boundary — specifically that the agent retains persistent access to real user folders (known folders) rather than scoped, ephemeral access. While crediting Microsoft for shipping it off by default and applying least-privilege thinking, the author argues key safeguards are missing: per-app scoping, tamper-evident audit logs, mandatory confirmations for destructive actions, and guaranteed on-device processing of private data. As an alternative, the author uses Terraform and Ansible to spin up isolated VMs connected to a local LLM, giving agents only the files they need. The conclusion is that OS-integrated agents are the right direction, but the current trust boundary is too loose for personal devices storing sensitive data.
