Microsoft open-sources Web Agent project: Magentic-UI! Make AI a truly “controllable, collaborative, and transparent” web execution assistant!

While developers around the world are still marveling at Copilot's code generation capabilities, Microsoft's AI Lab has released another open source bomb at the Build 2025 Developer Conference - MAgentic-UI!

This is the new generation of multi-agent Web operating system launched by Microsoft after Magentic-One!

It can:

It can automatically browse web pages, read files, write code, and work with you to complete tasks.

But it is not a fully automatic robot. It is a partner that you work with AI. You have the ultimate control and it is only responsible for doing the work.

• All operations will be displayed to you, and your consent will be sought in advance when clicking a button, opening a page, or sending a message.

• When executing a task, you can pause, modify the plan, and restart the process at any time.

• You can also set which actions require manual approval.

• Automatically click and fill in forms when browsing web pages

• Write code and return the running results

• Read the document and answer the questions

• Users can set task steps and modify and pause them at any time

Collaborative planning: Before the task begins, Magentic-UI will generate a detailed step-by-step plan. Users can adjust and confirm the plan according to their needs and expectations to ensure that the task proceeds as expected.

Collaborative execution: During the task execution process, the system will display the actions to be performed in real time, and the user can take over control at any time to ensure that the task execution meets the requirements.

Safety mechanism: Before performing important or irreversible operations, Magentic-UI will ask for the user's consent and support user-defined approval rules to protect operational safety. This effectively avoids the losses that may be caused by AI decision-making errors, making users feel more at ease during use.

Learning reuse: After the task is completed, the system will save the execution plan, and users can reuse or modify these plans in future tasks to improve work efficiency. As the number of uses increases, Magentic-UI will understand the user's work habits and needs more and more, and provide more personalized and efficient services.

• The Neural Style Engine (NSE) uses a GPT-4-like diffusion model that can dynamically compile natural language descriptions (such as "technological dark blue gradient") into CSS variables while ensuring multi-browser compatibility.

• Component DNA map Each component has a traceable version gene, which supports intelligent component reuse across projects. A test in an open source community showed that the component duplication rate was reduced by 92%.

• Real-time performance prediction: When dragging and dropping layouts, AI predicts page loading performance and gives optimization suggestions. Early adopters report that the first screen loading time is reduced by an average of 40%.

  
  

  
  

While we were amazed at the magic of Magentic-UI, Chris Novak, head of Microsoft's AI lab, said: "This is just the first stage in the evolution of intelligent development tools." 

We expect it to play a greater role in future digital life and work, bringing us a more efficient, intelligent and safe experience.

GitHub: https://github.com/microsoft/magentic-ui