Content area
The four other sub-agents inside Magnetic-One include WebSurfer, FileSurfer, Coder, and Computer Terminal. The Computer Terminal agent, meanwhile, gives the multi-agent system access to a console shell where Coder’s programs can be executed and new programming libraries can be installed. Microsoft warned potential users of Magnetic-One to take precautions, including running the agents in containers to isolate them and prevent any direct attacks, and monitoring logs closely to detect and mitigate risky behavior of agents. [...]it warned developers against giving the agents access to sensitive data or any resources that stand a chance of being compromised.
Microsoft wants enterprises to believe that its Magnetic-One multi-AI agent system will enable them to automate complex tasks that previously required human intervention.
One of a number of Agentic AI offerings to arrive on the market in recent months, Magnetic-One is built on Microsoft’s previously released AutoGen open-source agent development framework.
Microsoft expects the generalist multi-agent system, which is also open, to be used for open-ended web and file-based tasks for now, but it aims to create a multi-agent system than can handle complex tasks involving reasoning, such as automatically ordering in food or arranging a delivery of a product.
The new multi-agent system will go some way toward answering Marc Benioff’s criticism of Microsoft’s existing AI offering: The Salesforce CEO has said, “Copilot is more like Clippy 2.0,” referring to the ill-fated and irritating animated paperclip that once offered assistance on writing a letter in Microsoft Office.
Magnetic-One has a multi-agent architecture in which one agent, the Orchestrator, directs four other agents to solve a given task.
“The Orchestrator plans, tracks progress, and re-plans to recover from errors, while directing specialized agents to perform tasks like operating a web browser, navigating local files, or writing and executing Python code,” the company explained in a blog post.
From Microsoft’s explanation, the Orchestrator agent sounds very similar to Salesforce’s Atlas reasoning engine which controls the agentic loop in Agentforce.
The four other sub-agents inside Magnetic-One include WebSurfer, FileSurfer, Coder, and Computer Terminal.
WebSurfer is an LLM-based agent that can manage tasks on chrome-based web browsers. FileSurfer can command a markdown-based file preview application in order to read local files of most types. Coder specializes in writing code, collecting information from other agents, analyzing it, and creating new artifacts. The Computer Terminal agent, meanwhile, gives the multi-agent system access to a console shell where Coder’s programs can be executed and new programming libraries can be installed.
Given the architecture and the sub-agents, it seems that Magnetic-One should be capable of completing tasks in a computer, akin to computer use capability that Anthropic is currently showing in a beta version — with the limitation that Magnetic-One’s computer access is restricted to browsing the web and viewing files.
The Magnetic-One system, at least in Microsoft’s current implementation, uses GPT-4o as the underlying LLM for all agents, but Microsoft said that any other LLM can be used for the various agents, including the Orchestrator, in combination with small language models that specialize in certain tasks for other agents.
However, the company pointed out that enterprises or developers should use an LLM with strong reasoning capabilities to get the most out of the multi-agent system.
It has also released an agentic evaluation tool, AutoGenBench, to help developers test any Magnetic-One implementation with the help of benchmarks.
Microsoft warned potential users of Magnetic-One to take precautions, including running the agents in containers to isolate them and prevent any direct attacks, and monitoring logs closely to detect and mitigate risky behavior of agents.
Additionally, Microsoft said, agents’ access to the internet should be limited, and a human user should always be present to supervise them.
Finally, it warned developers against giving the agents access to sensitive data or any resources that stand a chance of being compromised.
Copyright Foundry Nov 6, 2024