Overview

Built on AutoGen version 0.4, Magentic-One assembles an Orchestrator, Coder, ComputerTerminal, FileSurfer, and WebSurfer for open-ended tasks across files, web pages, code, and computer actions. Its key control mechanism is two ledgers: a task ledger for facts, lookups, derivations, guesses, and plan, and a progress ledger for completion, repetition, progress, next speaker, and instruction. A stall threshold triggers reflection, plan revision, and context reset to support recovery, while tool boundaries and limited cross-task memory remain real constraints.

Learning outcomes

Magentic-One is an application architecture on AutoGen 0.4, not the definition of the AutoGen framework.
The task ledger preserves long-horizon facts and plans; the progress ledger diagnoses completion, repetition, progress, and the next assignment.
When stalls exceed the reported threshold of 2, Magentic-One revises the plan and clears contexts, trading stale assumptions for possible loss of useful local state.

Keywords

AutoGenMagentic-OneOrchestratorledgersplanningrecovery

From Framework to Generalist Team: Magentic-One

Magentic-One is a concrete test of how AutoGen’s conversation-centered design can support open-ended work rather than a single narrow application. The important question is not whether its five-agent roster should be copied, but how the system turns specialized capabilities, shared task state, and progress signals into a recoverable workflow. Understanding that architecture lets the reader distinguish a framework substrate from a particular generalist team and see why planning and replanning matter when a conversation spans tools, files, web pages, and code.

A framework becomes a team by assigning tools and boundaries

The 2024 Magentic-One paper describes a generalist multi-agent system implemented on AutoGen version 0.4. Its team contains an Orchestrator, a Coder, a ComputerTerminal, a FileSurfer, and a WebSurfer. The names are useful only if they are read as responsibility boundaries. The Orchestrator directs the work. The Coder writes and reasons about programs. The ComputerTerminal executes code or shell commands. The FileSurfer handles local files through a read-only preview application. The WebSurfer operates a browser and reports what it observes.

This is not a claim that AutoGen itself consists of those five agents. AutoGen is the framework that provides the conversational substrate and extensibility. Magentic-One is an application architecture built on that substrate for a particular scope: ad hoc tasks involving files, web interaction, coding, and computer actions. The distinction matters because a framework can support many topologies, while an application must choose a concrete team, control policy, tool set, and termination behavior. Treating one reference team as the definition of the framework would confuse an example of composition with the space of possible compositions.

Magentic-One uses a tool-centric division of labor. Instead of giving each participant a human-style title such as researcher, analyst, or critic and asking each to operate whatever tools are needed, it assigns broad capabilities to agents that are centered on a web browser, files, code writing, or code execution. The Orchestrator chooses the broad capability. The selected worker then chooses among a smaller set of actions within that capability. The hierarchy reduces the immediate decision space: the Orchestrator need not choose among every browser click, file operation, or line of code at once.

The WebSurfer illustrates the boundary. It receives a natural-language request, maps it to a browser action, and reports the new page state with a screenshot and written description. Its action space includes navigation, page interaction, and reading tasks. The Orchestrator does not directly manipulate every page element; it relays a task to a participant that understands the browser state. Similarly, FileSurfer navigates a read-only preview application and can work across several file types, while the ComputerTerminal carries out execution deterministically when asked. The agents are different not because they have different personalities, but because the outside state and action vocabulary differ.

The Coder occupies another boundary.

Published by DeepMox

FAQ

What agents are in the Magentic-One team built on AutoGen?

The Magentic-One team built on AutoGen 0.4 contains an Orchestrator, Coder, ComputerTerminal, FileSurfer, and WebSurfer. Each is centered on a different responsibility or tool boundary, rather than representing five agents that define AutoGen itself.

What is the difference between Magentic-One’s task and progress ledgers?

Magentic-One’s task ledger records facts, lookups, derivations, educated guesses, and the plan; its progress ledger tracks completion, looping, forward progress, the next speaker, and the next instruction. The first maps the problem, while the second diagnoses the current route.

How does Magentic-One recover after a stalled plan?

After the reported stall threshold of 2 is exceeded, Magentic-One leaves the inner loop, reflects on the failure, updates the task ledger, revises the plan, clears agent contexts, and starts a new inner cycle. This can remove stale assumptions but may discard local context not preserved in the ledger.

Previous10 minAutoGen Dynamic Group Routing

Next12 minAutoGen Magentic-One Evaluation

AutoGen Magentic-One Architecture

From Framework to Generalist Team: Magentic-One

A framework becomes a team by assigning tools and boundaries

FAQ