Name: Internet of Agents
Uploaded: 2024-12-20T05:59:00.718Z
Duration: 2 min 33 s
Description: An open-source framework inspired by the Internet, designed to facilitate collaboration among heterogeneous AI agents across diverse environments.

internet-of-agents

An open-source framework inspired by the Internet, designed to facilitate collaboration among heterogeneous AI agents across diverse environments.

Internet of Agents

Can we build an internet like system for autonomous agents? This was a question in the minds of these researchers at Beijing University who have built the Internet of Agents or iosa system for large language models inspired by how the internet works. More specifically, decentralized collaborative projects such as Wikipedia and Linux. Just like how the Internet connects people, a system like IA can connect different AI agents across different environments where agents can form teams and sub teams on their own and adapt to complex tasks.

Originally discussed in this paper, Internet of Agents weeing a web of heterogeneous agents for collaborative intelligence, the Ia system proposes a scalable platform for llm based multi-agent collaboration. The idea is to use autonomous agents that are capable of achieving near human performance on a wide range of tasks and get them to break break down complex task into.

Executable steps by using various tools and learning from feedback and experience. As with other area agents, this agentic framework also revolves around multitasking using the asynchronous task execution capabilities in this system. As this conceptual layered architecture shows, the Ia system can bring together agents with different skills and backgrounds in what they call heterogeneous agent integration. This is a really interesting feature as it combines custom agents with third party agents. 

The Ia framework consists of two main components: the server and the client. The server acts as a central hub managing agent registration, discovery, and message routing. The most interesting part of it is that it enables agents with varying capabilities to find each other and initiate communication. The client, on the other hand, serves as a wrapper for individual agents, providing them with the necessary communication.

Functionalities these two server and client components sit on a layered architecture with three layers of interaction. Data and foundation on the server side, the interaction layer facilitates team information and agent communication. The agent query blog enables agents to search for other agents based on specific characteristics. The group setup block facilitates the creation and management of group chats, and the message routing blog ensures efficient and accurate routing of messages between agents and group chats. 

The data layer manages information related to agents, group chats, and tasks. In this layer, the agent registry block maintains a comprehensive database of registered agents, including their capabilities and current status. The session management block manages active connections and ensures continuous communication between the server and connected clients. The final layer or the foundation.

Layer manages the entire system by providing essential infrastructure for agent integration, data management, and network communication. In this layer, the data block is responsible for data persistence and retrieval. The network block manages network communications, and the security block implements authentication, authorization, and other security protocols. 

The client side also has a super interesting and thoughtful design. The interaction layer manages the agent interactions within the system. The team formation blog implements the logic for identifying suitable collaborators and forming teams for the task at hand, similar to coalition formation in conventional multi-agent research. Complementing this, the communication blog manages the agent participation in group chats and handles message processing. 

The data layer functions as the agent's memory and maintains local data related to the agent's operations. It suggests sharing the AI Agent Store with colleagues.

Includes the agent contact block for storing information about other agents the current agent has interacted with. The group info blog for maintaining details about ongoing group chats and collaborations. And the task management blog for tracking the status and progress of tasks assigned to the agent. 

And finally, the Father Figure Foundation layer provides the basic functionalities for the client's operations. The agent integration blog defines the protocols and interface basis for integrating third party agents into IA ecosystem. Alongside this, the data blog handles local data storage and retrieval while the network block manages network communications with the server. 

I'm sure you agree that this is a very thoughtful architecture considering all practical aspects of running the system and bringing together in-house and thirdparty language models. The authors have also put a lot of thought and.

Consideration into the flow of conversation among agents. As you can see in this image, you can easily see how this is implemented through this five State Design. In the discussion State, the agents engage in general dialogue, exchanging ideas and clarifying task requirements. 

In the synchronous and asynchronous task assignment, tasks are assigned to specific agents with and without pause. The pause and Trigger State waits for the completion of the tasks in the previous state. Finally, the conclusion state produces a final response. 

They also show how this system works in action when producing your research paper, as you see in this image. The interaction between these four agents of AI researcher, Google Scholar agent, academic writer agent, and PDF processing agent is evident. As you can see, IA also outperformed both single LLM such as GPT 4 as well as other agentic systems such as Autogen across General AI.

Assistance benchmark even in the more more challenging level two and level three tasks which demand advanced reasoning and intricate collaboration. The IA system also outperform both autog gbt and open interpreter in various testing categories such as math coding and search and report which shows IA's proficiency in gathering and synthesizing information as well as in collaborative problem solving. 

Feel free to postuse the video to read these example instructions across the four mentioned categories that included in the IA open-ended benchmark. But that's not all IA also outperform multi-root collaboration as you see in the visualization of Rocko bench which is a tabletop test of manipulation of objects that are easy to understand for LLMS but hard to collaborate. 

For example, the cabinet task requires three agents to collaborate. Two agents open and hold the cabinet door while the third agent.

Retrieves two cops from inside the cabinet and places the onto coasters that match the color of the cups. I will leave a link so you can check out other fun robot tasks such as the famous sandwich task, the challenging sword task, and the frustrating rope task for multi-agent collaboration systems such as IA. If you want to know how this system is different than the mixture of Agents or Mo MOA system, check out this video of M Mo explanation and how to run a Cod.

Note, that there may be slight differences in text and video. Watch video for the original content.

Internet of Agents Video