Magentic-UI

Magentic-UI is a research platform for human-AI collaboration that pairs an on-device orchestrator model with a specialized browser-use model to deliver strong agentic performance without frontier-scale compute. It introduces co-planning and co-tasking, action guards for sensitive operations, task learning across sessions, and Sentinel Steps for long-running monitoring tasks. Browser sessions execute inside a lightweight VM sandbox (Quicksand) so that agent actions are isolated from the user’s environment, and the UI keeps users informed and in control through transparency mechanisms that allow steering, approval, or full takeover at any point.

Magentic-UI closes the gap between research demos and practical agentic tools by showing that sophisticated automation — web research, form filling, file management — can run efficiently with smaller models on consumer hardware while still preserving safety guarantees. Deliberate checkpoints for high-impact actions and approval workflows give users meaningful control, and iterative refinement against real user feedback informs the platform’s evolving patterns for human-agent collaboration. It is the production-leaning sibling of Magentic-One in Microsoft’s agentic research line.