← Back to Innovations
Agent System Embodied & GUI Experimental

Magentic-One

Generalist Multi-Agent System

Explore Magentic-One on GitHub → Explore on GitHub →
Magentic-One

About Magentic-One

Magentic-One is a generalist multi-agent system for solving open-ended web and file-based tasks. A lead Orchestrator coordinates specialized agents — a WebSurfer for browser navigation, a FileSurfer for local file management, a Coder for code analysis and generation, and a ComputerTerminal for code execution — and maintains a Task Ledger and Progress Ledger to plan, decompose work, and recover from errors. The Orchestrator runs a two-loop control structure: an outer loop revises the Task Ledger when work stalls, and an inner loop assigns subtasks while tracking progress.

Built on AutoGen, Magentic-One reaches competitive results on challenging benchmarks (GAIA, AssistantBench, WebArena) without per-task tuning, demonstrating that a small, modular cast of specialized agents can match heavier monolithic approaches on complex multi-step work. The plug-and-play architecture makes it straightforward for researchers and developers to add or replace agents for new domains, which has made Magentic-One a reference design for generalist agentic systems and the conceptual ancestor of both Magentic-UI and Magentic Marketplace.

Key capabilities

  • Orchestrator + specialized agents with Task/Progress Ledgers
  • Adaptive planning for complex web and file-based tasks
  • WebSurfer, FileSurfer, Coder, and Terminal specialized agents
  • Generalist system across browser, code, and file workflows
  • AutoGen-based open-source Python implementation
Technology Stack
Python AutoGen
Technology Stack
Python AutoGen