Magentic-One | An instantiation of thinker/actor pattern...

Multi-Agent systems are evolving to not only process information, but also act on it with human supervision. Magentic-One, a generalist multi-agent system, embraces this evolution through a pattern, which I call the Thinker-Actor pattern, combining reasoning and execution to manage complex workflows. I have recently started to explore this framework which seems not yet ready for production. Unlike controlled flow engineering approach, this framework takes more of an autonomous approach to orchestrate the agents.

What Is the Thinker-Actor Pattern?

The Thinker-Actor pattern in a multi agent system divides responsibilities between two roles:

1. Thinkers: Responsible for reasoning, planning, and decision-making. In Magentic-One, the Orchestrator agent exemplifies this role, breaking down tasks, tracking progress, and adapting strategies based on outcomes.
2. Actors: Execute specific actions based on the Thinker’s instructions. Magentic-One employs specialized agents—such as WebSurfer for web navigation, FileSurfer for file handling, Coder for code creation, and ComputerTerminal for execution.

This separation ensures clear accountability, modularity, and a streamlined workflow, making it easier to address both straightforward and complex tasks.

How Magentic-One Operates

Magentic-One’s Orchestrator acts as the central Thinker, creating plans and coordinating the efforts of its Actor agents. For example, when tasked with generating a report on recent AI safety papers, the system breaks the problem into steps:

1. Thinker Role: The Orchestrator plans the sequence of actions, such as searching for relevant papers, downloading them, and summarizing key points.
2. Actor Roles: By iterating between planning and execution, Magentic-One ensures the task is completed effectively, even when errors or unexpected challenges arise.

Advantages of the Thinker-Actor Pattern

This division of responsibilities offers several benefits:

• Clarity and Modularity: Each agent has a specific role, simplifying development and enabling easy integration of new capabilities.
• Resilience to Errors: The Orchestrator continually monitors progress and adapts plans, ensuring tasks can recover from mistakes.
• Specialized Focus: Actors excel in their domains, reducing cognitive load and improving performance for each task.

This framework have been open sourced and available at https://github.com/microsoft/autogen/tree/main/python/packages/autogen-magentic-one

Other references:

https://www.microsoft.com/en-us/research/uploads/prod/2024/11/MagenticOne.pdf