Automate your web tasks while you stay in control
Magentic-UI by Microsoft Research is an open-source human-centered web agent. Collaboratively plan & execute web tasks with AI, featuring co-planning, action guards & plan learning. Built on AutoGen.
Magentic-UI's multi-agent architecture includes WebSurfer for browser control, Coder for Python and shell execution, and FileSurfer for file conversion and analysis. This allows users to combine web automation, data processing, and document handling in one coordinated task.
Magentic-UI enables co-planning and co-tasking, allowing users to review and modify the agent's step-by-step plan before execution, pause mid-task to provide feedback, and take direct browser control when needed. This human-in-the-loop approach ensures users understand what the agent is doing and can intervene if something goes wrong.
Magentic-UI includes action guards that require user approval before executing consequential or irreversible actions (e.g., closing tabs, submitting forms, making payments). Users can configure approval frequency, balancing automation with safety and control over critical operations.
Magentic-UI's plan learning feature allows users to save step-by-step plans from successful task completions and reuse or modify them for future similar tasks. Users can view and edit saved plans in a gallery, reducing setup time for repetitive work patterns.
Magentic-UI is an open-source research prototype designed specifically to study human-in-the-loop approaches and oversight mechanisms. Researchers can use it to investigate co-planning, co-tasking, action guards, and plan learning as design patterns for trustworthy AI agents.