Microsoft Magentic-One

About

Magentic-One came out of Microsoft Research as a demonstration that a small set of generalist agents, well-coordinated, could match specialized systems on open-ended tasks — GAIA, AssistantBench, and WebArena — without task-specific tuning. The Orchestrator keeps a Task Ledger of facts, guesses, and the current plan in an outer loop, re-planning when progress stalls, and an inner Progress Ledger that each step assigns the next subtask to one of four specialists: a Chromium-driving WebSurfer using accessibility-tree and set-of-marks prompting, a FileSurfer reading documents through markdown previews, a Coder, and a ComputerTerminal. Specialists are interchangeable — the system keeps working when one is swapped or removed — and per-agent model assignments allow heterogeneous LLM configurations. It ships as part of autogen-agentchat (MagenticOneGroupChat) with AutoGenBench for isolated, repeated benchmark runs, and Microsoft positions it for research use inside sandboxed containers with human monitoring rather than as a production harness.