Magentic-One

Magentic-One is a generalist agentic system published by Microsoft Research that addresses the challenge of completing open-ended, multi-step tasks that require coordination across heterogeneous capabilities. The system's architecture centers on an Orchestrator agent that maintains an overall task plan, tracks progress against that plan, and dynamically re-plans when an agent encounters an error or an unexpected result. The Orchestrator directs four specialized agents: WebSurfer, which controls a Chromium browser using Playwright to navigate websites, interact with web elements, and fill out forms; FileSurfer, which browses local file systems, reads documents, and navigates directories; Coder, which writes Python code to solve analytical or transformation problems; and ComputerTerminal, which executes code and performs system-level operations. The system is model-agnostic and has been tested with GPT-4o as the primary model and OpenAI o1 for the outer Orchestrator reasoning loop. Magentic-One is available as a package within the AutoGen AgentChat framework, installable via pip, and has since been ported to use the autogen-agentchat interface for a more modular implementation. Microsoft also released AutoGenBench alongside Magentic-One, a standalone agentic evaluation tool that supports rigorous isolated benchmark execution. AI researchers use Magentic-One as a reference implementation and baseline for studying multi-agent coordination, error recovery, and task planning in agentic systems. Developers building enterprise automation tools use the modular agent architecture as a foundation for adding or removing specialized agents without requiring full system restructuring—an advantage the research team explicitly compares to object-oriented programming's encapsulation principle. Explore this category. Data scientists extend the Coder and ComputerTerminal agents for automated data pipeline execution, scientific computing workflows, and analytical reporting. The WebSurfer agent has been used in research contexts for automated form completion, web-based data extraction, and web navigation benchmarking. An evolved version of Magentic-One's architecture forms the foundation of Magentic-UI, a human-in-the-loop web agent released by Microsoft Research in 2025. Magentic-One's key differentiators are its benchmark-validated multi-agent coordination architecture, its plug-and-play agent extensibility, and its full open-source availability under the AutoGen ecosystem. Setup requires Python, Docker containers for safe isolated execution (strongly recommended by Microsoft), and an LLM API key for GPT-4o or a compatible model. Microsoft explicitly acknowledges in the project documentation that Magentic-One operates in digital environments designed for humans and carries inherent risks: agents may take undesirable actions, accept cookie agreements autonomously, or be susceptible to prompt injection from web pages. Human oversight and containerized execution are recommended requirements for any responsible deployment. The system is described as a research implementation rather than a production-hardened tool, and Microsoft continues to develop it alongside Magentic-UI for supervised use cases Explore more.