Magentic One: Microsoft’s Revolutionary Multi-Agent AI System

Key Takeaways

Microsoft has just introduced something really exciting called magentic 1 a powerful multi-agent AI system that's set to change how we get things done I know multi-agent AI might sound a bit complex but think of it as a team of AI experts each with a specific skill working together step by step to handle all kinds of tasks across different areas unlike the usual AI systems that just give you answers magentic 1 goes beyond that whether it's doing research on various web pages navigating files on your computer or even ordering your lunch this system handles it all seamlessly it can also browse the web edit documents and automate complex tasks effortlessly let's dive into to see how these agents work and what makes magentic one so powerful we start on the autogen magnetic 1's GitHub page where we see a highlevel overview of how magnetic 1 actually works the task begins with the orchestrator which breaks it into subtasks and assigns them to specialized agents coordinating progress step by step the file Surfer agent handles files extracting and processing data like code or documents coder analyzes writes and refines code while computer terminal executes it in a controlled environment web Surfer navigates and interacts with web pages completing actions like retrieving information or extracting content together these agents collaborate to efficiently solve complex open-ended problems the orchestrator manages the task by creating and updating a task Ledger which includes facts computations and a task plan it monitors progress through the progress Ledger identifying if the task is complete or if adjustments are needed the stall count section represents a checkpoint in the orchestrator workflow for monitoring progress if the syst system detects that progress is not being made it increases the stall count when the stall count exceeds a threshold the orchestrator determines that the current approach may not be effective and revisits the task ledger to update or revise the plan this mechanism ensures that the system doesn't get stuck in unproductive loops and dynamically adjusts to achieve the task more effectively this iterative process continues until the task is successfully completed let's try it out to get a for it we first copy the Clone command we paste the command into a new directory and see that autogen is cloned let's continue with the second code block it first changes into the directory of autogen magnetic one and then installs all dependencies with Pip install to start the example script we can use different options for customization the logs deer option specifies where to save logs downloads and browser screenshots and is required for every run the hill mode option enables human in the loop mode allowing real-time supervision and intervention during execution for visual tracking the save screenshots option captures browser screenshots throughout the task let's choose the save screenshot option to visually track what's happening during the task execution let's paste the command and start start with our first objective time to do some research let's give it the task describe Trends in Germany regarding AI all right the agents go to work we see that it first considers the facts and interestingly it also builds a corresponding plan the plan consists of first asking the web server to research recent news and articles then looking for statistics regarding AI adoption in different Industries in Germany and finally summarizing the findings into a coherent overview we see how the orchestrator hands over to the web Surfer agent the web Surfer takes over and searches Bing for AI Trends in Germany interestingly it also takes a screenshot of the current step which we can look at here we see the Bing screenshot including the search term and how the individual components were analyzed in the next step we also see that automatic OCR scanning has occurred and the results from this search query are being evalu valuated accordingly we see the results here the orchestrator says to continue the search with a focus on government initiatives then it continues with further searching we see that it has found an article that apparently fits well there's a back and forth between the web server and the orchestrator here we see the website of the article that was then evaluated it's a bit like hiring a small team that works together in an office collaborating to to complete tasks for you and here is the final answer based on the search results and the article it found the agents have found out that Germany is making progress in AI with government plans industry efforts and research but it faces challenges like low adoption and policy gaps and needs better skills infrastructure and inclusivity