Magentic-One
Generalist multi-agent system that solves complex web and file tasks via an orchestrator-led team.
About
Magentic-One is a high-performance generalist multi-agent system developed by Microsoft Research, released in November 2024. It is built on AutoGen and features an Orchestrator agent that leads a team of specialized sub-agents: WebSurfer (browser automation via Chromium), FileSurfer (local file and document exploration), Coder (Python code synthesis), and ComputerTerminal (shell execution). The Orchestrator maintains a Task Ledger of facts, guesses, and plans plus a Progress Ledger for step-by-step delegation, re-planning when progress stalls.
It demonstrated competitive results on GAIA, AssistantBench, and WebArena without task-specific tuning, and now ships as MagenticOneGroupChat inside autogen-agentchat, with AutoGenBench for isolated benchmark runs. Microsoft positions it for research use inside sandboxed containers with human monitoring rather than as a production harness. It is open source and model-agnostic, supporting various LLMs. Limitation: it is a research demonstration system, not a hardened production product — expect to do integration work to deploy it for real workloads.
What it can do
- Orchestrator-led hierarchical multi-agent planning with task and progress ledgers
- WebSurfer browser automation and FileSurfer document exploration agents
- Coder and ComputerTerminal for code synthesis and sandboxed execution
- Benchmark-ready via AutoGenBench on GAIA, AssistantBench, and WebArena
Setup
Available as MagenticOneGroupChat in the autogen-agentchat package; runs in Docker containers. See the AutoGen repo's autogen-magentic-one package.
Prerequisites
Python environment, Docker for sandboxed execution, and an LLM API key (supports GPT-4o and other models).
Use cases
Works with
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 71; Trending score 82/100 from researcher assessment.
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.