-
Background The authors discuss how the rapid advancement of Large Language Models (LLMs) has enabled the creation of highly capable autonomous agents, but current multi-agent frameworks face challenges integrating diverse third-party agents, simulating distributed environments, and adapting to dynamic task requirements due to their limitations.
-
Existing Work The existing work has three main constraints: ecosystem isolation, single-device simulation, and rigid communication and coordination, impeding the integration of diversified third-party agents, simulation of distributed environments, and swift adaptation to dynamic tasks.
- Introducing the Internet of Agents (IoA) framework
-
Challenge 1: Integration of third-party agents and simulation of distributed environments Existing multi-agent frameworks struggle with integrating diversified third-party agents and supporting distributed environment simulation. IoA addresses this challenge by introducing an agent integration protocol and an instant messaging app-like architecture, enabling seamless integration and effective collaboration of diverse third-party agents running on different devices.
-
Challenge 2: Flexibility and automation in communication and collaboration The rigid communication process and state transitions of existing frameworks are limitations that IoA overcomes by proposing a flexible finite-state machine mechanism and dynamic teaming and conversation flow control components. This allows agents to autonomously decide their state based on the current task, effectively facilitating discussion and sub-task execution.
-
Extensive experiments on general assistant tasks, embodied AI tasks, and retrieval-augmented generation benchmarks demonstrate that IoA consistently outperforms state-of-the-art baselines. Notably, with the integration of AutoGPT and Open Interpreter, IoA achieved win rates of 66% to 76% in open-domain task evaluations, surpassing individual performances of these agents. In the field of RAG question-answering, IoA, with a GPT-3.5-based implementation, significantly surpassed existing methods and closely matched or even exceeded the performance of GPT-4.
The paper presents a flexible and scalable platform for multi-agent collaboration, the Internet of Agents (IoA), which overcomes the limitations of existing frameworks and demonstrates superior performance across multiple tasks and application scenarios. Furthermore, the release of the codebase facilitates further development in autonomous agent systems.