← All Papers Orchestrator Specialists Ledger Simulator GAIA Results Quiz
Microsoft Research · AutoGen · Nov 2024

Magentic-One: A Generalist
Multi-Agent System

How a centralized dual-loop Orchestrator directs specialized web, coding, and filesystem agents to conquer open-ended real-world multimodal tasks.

Inspect Dual Ledger Meet the Specialists
Authors: Adam Fourney, Gagan Bansal, et al.
Institution: Microsoft Research
Framework: Built on AutoGen
Key Feature: Dual-loop Ledger Orchestration + WebSurfer
Core Architecture

The Dual-Loop Orchestrator Pattern

Magentic-One places a centralized Orchestrator in charge of all subagent delegations, maintaining global coherence and dynamic replanning.

Outer Loop: Planning & Replanning

The outer loop maintains the Task Ledger. It formulates initial high-level sub-goals, tracks facts learned across tasks, and triggers wholesale strategy revisions if multiple subagents stall.

Stall Counter: If the inner loop takes more than 3 turns without measurable progress, the outer loop executes a reset, clearing hallucinated context.
Inner Loop: Progress Assessment

The inner loop maintains the Progress Ledger. After every single subagent tool call or message, it evaluates whether the current sub-step succeeded or failed.

Dynamic Dispatch: Selects the next specialist (WebSurfer vs Coder vs FileSurfer) based on immediate intermediate deliverables.
Agent Ecosystem

The Four Specialist Subagents

Each agent is purpose-built with tailored prompts, actions, and isolated sandboxes.

1. WebSurfer

Autonomous Browser

Operates Chromium via Playwright. Navigates, clicks interactive elements, scrolls pages, solves captchas, and extracts clean markdown DOMs.

2. FileSurfer

Document Explorer

Inspects local files, parses complex PDFs, parses multi-sheet Excel files, and runs OCR on scanned diagrams.

3. Coder

Algorithm Writer

Writes statistical Python scripts, data cleaning routines, and web scraping regexes tailored to subagent requests.

4. ComputerTerminal

Sandbox Executor

Executes Python and shell scripts inside an isolated container, capturing stdout, stderr, and returned figures.

Interactive Simulation

Magentic-One Ledger State Machine

Observe how the Orchestrator tracks Task and Progress Ledgers on a multimodal GAIA benchmark question.

TASK: "Find the 2023 revenue of Microsoft's cloud segment, write a script to plot year-over-year growth, and save the chart."
📋 TASK LEDGER (Outer Loop)
⚡ PROGRESS LEDGER (Inner Loop)
Benchmark Evaluation

State-of-the-Art on GAIA & AssistantBench

Evaluating performance on real-world multimodal problem solving.

GAIA Benchmark (General AI Assistants)
Single Agent (GPT-4o) 32.4%
OpenAI Assistants API + Tools 34.8%
Magentic-One (Multi-Agent) 47.8%
Why the Architecture Wins
  • Tool Specialization: WebSurfer handles browser state trees without polluting the Coder's namespace.
  • Stall Recovery: When WebSurfer hits an anti-bot wall, the Orchestrator pivots to FileSurfer searching downloaded local annual reports.
  • Stateless Agent Reset: Subagents can be rebooted fresh with clean context while the Orchestrator preserves the global goal.
Test Yourself

Magentic-One Mastery Quiz

Assess your understanding of orchestrator-led multi-agent architectures and specialized tool routing.