On Monday, Microsoft introduced a new agentic web interface capable of executing complex online tasks. The tool, dubbed Magentic-UI, was unveiled during the company’s Build 2025 conference and is described as a “human-centred web agent.” Microsoft emphasizes that, unlike traditional computer use agents, Magentic-UI actively collaborates with users to accomplish projects, providing transparency by displaying its actions in a dedicated panel. As an open-source initiative, the new interface is available on both GitHub and Microsoft’s Azure AI Foundry Labs.
Magentic-UI Launches as Microsoft’s First Computer Use Agent
In a blog post outlining its features, Microsoft detailed how the Magentic-UI functions. It represents a network of agents designed for computer-related tasks, comparable to offerings from OpenAI and Anthropic. These agents can navigate through a virtual machine and utilize applications like web browsers to undertake various actions, including reserving restaurant tables or making online purchases.
According to Microsoft, Magentic-UI can browse the internet, write and execute Python and shell code, and analyze files. The company highlighted its “human-centric” approach, allowing users to modify the agent’s plans, pause operations, and provide feedback in natural language, or regain control of their browser interface.
Magentic-UI’s interface
Photo Credit: Microsoft
Moreover, before executing any irreversible actions—such as placing a restaurant order or adding items to a shopping cart—the agent will seek user permission. At this point, users can also alter the commands being processed. Additionally, Magentic-UI has the capability to learn from past tasks, using this knowledge to enhance performance on future assignments.
The architecture behind Magentic-UI is built upon Microsoft’s Magentic-One agentic system. The tool comprises four distinct AI agents cooperating to fulfill tasks. Central to this system is an Orchestrator, which acts as the lead agent and collaborates with the user to formulate task plans. This component also processes feedback and designates sub-tasks to the other agents. The WebSurfer agent contains a web browser and is capable of clicking, typing, scrolling, and navigating through pages to accomplish tasks.
There is also a Coder agent that features a Docker code-execution container, capable of writing and executing Python and shell commands. Lastly, the FireSurfer agent, equipped with similar Docker capabilities, assists with file conversion processes.
Individuals interested in utilizing the AI agent locally can download it from GitHub, where Microsoft is offering its research preview under a permissive MIT license. For users preferring a cloud-based solution, Magentic-UI is also available for access through Microsoft’s Azure AI Foundry Labs.
