AutoGenMagenticOne
title: AutoGen Magentic-One type: platform created: 2026-05-26 last_updated: 2026-05-26 related: ["radar/tools/OpenHands", "radar/tools/BrowserUse", "radar/tools/Fabric"] sources: ["https://github.com/microsoft/autogen/tree/main/python/packages/autogen-magentic-one"] radar_quadrant: Platforms radar_ring: Assess radar_position: inner
Microsoft Research's generalised multi-agent system built on AutoGen. An Orchestrator agent holds the plan and delegates to four specialised sub-agents — WebSurfer (browser), FileSurfer (filesystem), Coder (code generation), and ComputerTerminal (sandboxed execution) — enabling complex multi-step tasks that span web research, file processing, coding, and command execution.
The system separates planning from execution across a fixed team of agents:
- Orchestrator — tracks the overall goal, maintains a task ledger, decides which sub-agent to invoke, detects stalls or errors, and replans
- WebSurfer — Playwright-based browser agent; navigates, clicks, fills forms, extracts content from any website
- FileSurfer — reads and navigates local files and directories; handles markdown, PDFs, code, and data files
- Coder — writes, reviews, and debugs code in any language
- ComputerTerminal — executes code and shell commands in a sandboxed Docker container; results feed back to the Orchestrator
The Orchestrator's separation from execution is the key architectural decision. Sub-agents do not need to track the broader task; they receive a specific instruction and return a result. This keeps each agent's context focused and reduces the goal-drift that single-agent systems exhibit on long tasks.
OpenHands (see radar/tools/OpenHands) is an autonomous coding agent optimised for software engineering workflows: reading codebases, implementing features, running tests. Magentic-One is more general-purpose — it handles tasks that cross capability domains (research a topic, process data, write a script, run it, summarise output). The two are complementary: OpenHands for development tasks, Magentic-One for cross-domain automation.
AutoGen Magentic-One sits in the Assess ring of the Platforms quadrant, at inner position. First studied via the AutoGen GitHub repository (2024-11-12). The Orchestrator + specialised sub-agents pattern addresses a real limitation of single-agent frameworks on multi-step tasks: context pollution and goal drift. Microsoft Research production use and active AutoGen ecosystem lower adoption risk. Inner position reflects sandboxed local execution (no cloud dependency), open-source availability, and applicability to research, data processing, and operational automation tasks beyond OpenHands' coding scope. Remaining gate before Trial is a completed multi-step task requiring at least three sub-agents with Orchestrator replanning observed at least once.
