This is the official implementation of the paper "AgentNet: Decentralized Evolutionary Coordination for LLM-Based Multi-Agent Systems" accepted by Neurips 2025.

AgentNet is a novel framework for building decentralized, privacy-preserving, and adaptive multi-agent systems (MAS) powered by large language models (LLMs). It addresses the limitations of traditional MAS architectures that rely on centralized controllers and static workflows.

📄 Paper Title: AgentNet: Decentralized Evolutionary Coordination for LLM-Based Multi-Agent Systems

👨‍🔬 Authors: Yingxuan Yang*, Huacan Chai*, Shuai Shao, Yuanyi Song, Siyuan Qi, Renting Rui, Weinan Zhang

🏫 Affiliation: Shanghai Jiao Tong University

📎 arXiv:2504.00587

This figure illustrates the overall architecture of AgentNet. It consists of multiple LLM-based agents connected in a dynamic, decentralized Directed Acyclic Graph (DAG). Each agent has its own retrieval-augmented memory, local routing strategy, and can evolve independently.

• 🧠 Fully Decentralized Architecture
  No central coordinator. Every agent makes autonomous routing and execution decisions, improving fault tolerance and enabling emergent collaboration.

• 🔀 Dynamically Evolving DAG Topology
  Agent connections adapt in real time based on task success metrics, forming a Directed Acyclic Graph (DAG) that optimizes collaboration.

• 📚 Retrieval-Augmented Adaptive Learning
  Agents store and retrieve relevant memory fragments from past tasks to refine their expertise over time, supporting continuous specialization.

This illustration compares conventional Pre-Defined Multi-Agent Systems (which are hierarchical, static, and prone to single points of failure) with AgentNet, which is fully decentralized, self-evolving, and dynamically specialized.

AgentNet enables fault-tolerant collaboration and adaptive skill growth without needing predefined roles or a central controller.

The following animation provides a comprehensive demonstration of AgentNet's decentralized multi-agent coordination:

• How tasks are dynamically routed, forwarded, split, and executed across agents
• How agents adapt and evolve their capabilities based on their past performance
• The emergence of specialized agent roles over time

📹 Watch Full Video Demo

If you use AgentNet in your research, please cite us as follows:

@misc{yang2025agentnetdecentralizedevolutionarycoordination,
  title={AgentNet: Decentralized Evolutionary Coordination for LLM-based Multi-Agent Systems}, 
  author={Yingxuan Yang and Huacan Chai and Shuai Shao and Yuanyi Song and Siyuan Qi and Renting Rui and Weinan Zhang},
  year={2025},
  eprint={2504.00587},
  archivePrefix={arXiv},
  primaryClass={cs.MA},
  url={https://arxiv.org/abs/2504.00587}
}