Microsoft tries to soothe fears as it prepares to test 'agent workspaces' (www.techradar.com)

Microsoft has revealed that Windows 11 will soon test "agent workspaces"—an experimental, developer-only feature surfaced in preview build 26220.7262 via a new "experimental agentic features" toggle in Settings. The private Windows Insider preview will let isolated AI agents run alongside a user but in separate, lightweight Windows accounts with scoped authorization and runtime isolation. Users can grant each agent explicit access to particular apps and files, configure permissions per agent, and review full activity logs. Microsoft says the design aims to balance autonomous task automation (examples include Copilot Actions for chores like de-duplicating photos and Manus AI for higher-level jobs such as building websites) with containment to limit lateral system access and resource impact. For the AI/ML community this is a significant step toward an "agentic OS" where persistent, multi-modal agents are first-class desktop entities—opening new UX and orchestration opportunities but also expanding the attack and failure surface. Key technical considerations include per-agent accounts, fine-grained file/app permissions, runtime isolation, telemetry/logging, and defenses Microsoft calls out (e.g., guarding against cross-prompt injection). The phased, limited rollout acknowledges risks—hallucinations, misconfigurations, and past issues (e.g., Recall) underscore that trust will hinge on robust sandboxing, secure defaults, and transparent auditing as agents move from preview to broader use.