microsoft/autogen

Magentic-One is Microsoft Research's open-source multi-agent orchestrator. Built on AutoGen. MIT licensed. Python. 8,500+ GitHub stars and climbing.It is the rare multi-agent framework that ships with the agents already built. You do not assemble specialists out of YAML and prompt fragments. You install Magentic-One and you get a working system on day one.The ArchitectureMagentic-One is built on the AutoGen v0.4 conversation framework, but it adds two things AutoGen left as exercises for the reader.First, an Orchestrator agent that owns the planning loop. It reads the task, builds a plan in a public ledger, assigns sub-tasks to specialists, monitors progress, and rewrites the plan when specialists fail or stall. The Orchestrator is the difference between a multi-agent demo and a multi-agent system that finishes.Second, four production-quality specialist agents:WebSurfer — operates a Chromium browser via Playwright. Navigates, clicks, fills forms, extracts content. Handles modern JavaScript-heavy sites, not just static HTML.FileSurfer — reads, navigates, and reasons about local files (PDFs, Word, Excel, CSV, plain text, code). Built on Microsoft's MarkItDown library.Coder — writes code, asks the Orchestrator to run it, reads stack traces, and iterates. Pure reasoning agent, no execution privileges of its own.ComputerTerminal — the executor. Runs shell commands and Python code in a controlled environment. The only specialist with side-effect privileges.The split between Coder (reasons about code) and ComputerTerminal (executes code) is the most important design choice in the system. It is what makes Magentic-One auditable in environments where letting an agent run arbitrary commands is a non-starter.What It Actually SolvesThe original Magentic-One paper benchmarked the system on three agentic tasks:GAIA — multi-tool reasoning on real-world questions. Magentic-One scored 38% — competitive with the best closed-source agent stacks at release.AssistantBench — open-web agentic tasks. Strong performance because WebSurfer is genuinely good.WebArena — simulated multi-website interactions. Solid but behind specialized web agents.The benchmark numbers matter less than the architecture. The system gets to those scores with four general-purpose specialists, not 40 hand-tuned agents. That is what makes it a practical starting point for production work.How to Use Itpip install magentic-one from magentic_one import MagenticOne import asyncio async def main(): m1 = MagenticOne() result = await m1.run( "Find the three most-cited papers from NeurIPS 2025 on agentic coding, summarize each, and save them to a Markdown file." ) print(result) asyncio.run(main())That single command kicks off the Orchestrator, WebSurfer searches arxiv and Google Scholar, FileSurfer parses the PDFs, Coder writes the summary code, and ComputerTerminal runs it. You can swap the model backend (default GPT-4o, easy to point at GPT-5.5 or Claude Opus 4.6) in the config.Where It Falls ShortTwo honest limitations. First, Magentic-One is not a low-code platform. You need real Python skills to extend the specialists or wire in custom tools. The drag-and-drop crowd should look at Microsoft Copilot Studio or Google's Gemini Enterprise Agents instead.Second, ComputerTerminal grants the agent shell execution. That is intentional and the docs are clear about it, but you must run the system in a sandboxed environment (Docker, devcontainer, dedicated VM) for any task involving untrusted inputs. The framework gives you the levers; the security configuration is on you.Who Should Install ItInstall Magentic-One if you are an engineering team building a multi-agent system from code, you want a strong orchestration baseline without writing one from scratch, and you are comfortable with Python and an LLM SDK. It saves weeks of glue work.Skip it if you want a hosted enterprise agent platform — Gemini Enterprise Agents and Copilot Studio are the right tier for that need.Related ResourcesArticle: GPT-5.5 just launched — the new flagship model you can plug into Magentic-One's Orchestrator and Coder.Tool: Gemini Enterprise Agents — the hosted alternative if you do not want to run a Python framework yourself.MCP server: Anthropic Claude Code MCP — wire into Magentic-One specialists to extend tool coverage.Skill: Anthropic Data Analysis Skills — structured skills the Coder agent can consume during analysis tasks.