Microsoft Research has unveiled Magentic-UI, an experimental open-source prototype of a human-centered web agent focused on assisting users with complex, action-oriented tasks through real-time collaboration in a web browser. Unlike fully autonomous agents, Magentic-UI prioritizes user involvement, offering granular control and oversight via features like collaborative planning and execution, real-time feedback, and configurable safety mechanisms. The agent is built on technologies including Magentic-One and the AutoGen agent framework, and is licensed under MIT, making it freely available for research and development.
Magentic-UI´s core features include co-planning, where users can modify the agent´s step-by-step plans before execution; co-tasking, which enables users to pause, intervene, or guide the agent during task completion; and action guards that prompt for explicit user approval before any potentially irreversible operations. The system also supports plan learning, allowing both the agent and users to save successful strategies for future automation or reference. Magentic-UI is integrated with Azure AI Foundry Labs and demonstrates strong modular architecture, with dedicated agents for orchestrating tasks, web browsing, code execution, and file management, all coordinated by a lead orchestrator agent.
Preliminary benchmarking using the GAIA standard revealed that incorporating simulated users (representing human expertise or extra task-specific knowledge) significantly improves the agent´s accuracy and task-completion rates over autonomous modes, demonstrating the benefit of human-in-the-loop interaction. Safety and security are ensured through features like allow-lists, Docker sandboxing, and layered permission checks, with the agent refusing or deferring risky actions when necessary. Magentic-UI invites further research into agentic safety, transparency, and human-agent cooperation, and Microsoft Research encourages the academic community to expand upon the platform to explore open questions about safeguarding, personalization, and effective collaboration within Artificial Intelligence systems.
