![](https://www.aicritique.org/us/wp-content/uploads/2024/11/microsoft.jpg)
On November 4, Microsoft announced “Magentic-One,” an AI system composed of five agents. This system can plan and execute complex multi-step tasks, such as writing code or ordering food, offering capabilities beyond traditional generative AI chatbots.
Magentic-One consists of the following five agents:
- Orchestrator: The lead agent that divides tasks and manages progress.
- WebSurfer: An agent capable of operating a browser.
- FileSurfer: Capable of reading local files.
- Coder: Performs code creation and information analysis.
- ComputerTerminal: Provides access to a console shell.
The Orchestrator creates a “Task Ledger” to plan tasks and uses a “Progress Ledger” to confirm task completion and track progress. In tests using the evaluation tool AutoGenBench, Magentic-One showed results surpassing GPT-4 but did not reach human-level accuracy. Microsoft warns that Magentic-One operates in a digital world designed for humans and may cause irreversible consequences.