Source-linked AI summary
AgentNet: Decentralized Evolutionary Coordination for LLM-based Multi-Agent Systems
Yingxuan Yang, Huacan Chai, Shuai Shao, Yuanyi Song, Siyuan Qi, Renting Rui, Weinan Zhang
TL;DR
Existing multi-agent systems often depend on centralized coordination, limiting scalability, adaptability, fault tolerance, and privacy-preserving collaboration. AgentNet uses decentralized DAG-based routing, dynamic agent connections, and retrieval-based memory for evolving specialization, and experiments report improved task efficiency, adaptability, and specialization. The paper also identifies routing in large, heterogeneous systems as an unresolved challenge.
Problem
Existing LLM-based multi-agent systems rely on centralized coordination or static workflows, creating scalability, adaptability, failure-tolerance, and cross-organizational privacy challenges.
Method
AgentNet distributes routing and execution across autonomous agents in a dynamically reconfigurable DAG, using retrieval-based memory to refine agent expertise and specialization.
Results
AgentNet improves task efficiency, specialization stability, and adaptive learning speed over traditional LLM-based multi-agent frameworks in dynamic environments.
Takeaways & Limitations
AgentNet provides a decentralized framework for scalable, fault-tolerant, adaptive, and privacy-preserving multi-agent collaboration.
Takeaways & Limitations
Routing remains challenging in large, heterogeneous systems because the router must identify suitable agents among many candidates with differing strengths and weaknesses.
Abstract
from arXiv · showhide
The rapid advancement of large language models (LLMs) has enabled the development of multi-agent systems where multiple LLM-based agents collaborate on complex tasks. However, existing systems often rely on centralized coordination, leading to scalability bottlenecks, reduced adaptability, and single points of failure. Privacy and proprietary knowledge concerns further hinder cross-organizational collaboration, resulting in siloed expertise. We propose AgentNet, a decentralized, Retrieval-Augmented Generation (RAG)-based framework that enables LLM-based agents to specialize, evolve, and collaborate autonomously in a dynamically structured Directed Acyclic Graph (DAG). Unlike prior approaches with static roles or centralized control, AgentNet allows agents to adjust connectivity and route tasks based on local expertise and context. AgentNet introduces three key innovations: (1) a fully decentralized coordination mechanism that eliminates the need for a central orchestrator, enhancing robustness and emergent intelligence; (2) dynamic agent graph topology that adapts in real time to task demands, ensuring scalability and resilience; and (3) a retrieval-based memory system for agents that supports continual skill refinement and specialization. By minimizing centralized control and data exchange, AgentNet enables fault-tolerant, privacy-preserving collaboration across organizations. Experiments show that AgentNet achieves higher task accuracy than both single-agent and centralized multi-agent baselines.
1 Introduction
AgentNet addresses scalability, adaptability, failure-tolerance, and privacy problems in centralized multi-agent systems through decentralized coordination, dynamic routing, and retrieval-based expertise refinement. Experiments report improved performance in dynamic environments.
- Motivation: Centralized or static multi-agent workflows constrain scalability, create single points of failure, hinder adaptation, and complicate cross-organizational collaboration because of privacy and proprietary-knowledge concerns.Rigid roles also limit dynamic reassignment and flexible use of agent expertise.
- AgentNet: AgentNet removes the central orchestrator and lets agents dynamically reconfigure connections and redistribute tasks through a Directed Acyclic Graph.The architecture is designed to support self-organizing, fault-tolerant collaboration while preventing cyclic dependencies.
- AgentNet: Its retrieval-based memory stores successful task trajectories, retrieves relevant examples through few-shot learning, and prunes less pertinent trajectories to refine expertise over time.This mechanism supports specialization while limiting memory growth.
- Privacy and robustness: Decision-making is distributed across agents, while local storage and minimal task-relevant metadata exchange reduce exposure of sensitive information.Dynamic topology confines data flow to necessary interactions, and retrieval memory prunes outdated trajectories.
- Results: AgentNet significantly outperforms traditional LLM-based multi-agent frameworks in dynamic environments, improving task efficiency, specialization stability, and adaptive learning speed.The evaluation identifies decentralized evolutionary coordination as effective in large-scale AI ecosystems.
2 Related Work
Prior work improves agent workflows, robustness, topology, prompts, or role specialization, but generally retains centralized control or focuses on individual adaptation. AgentNet instead combines decentralized coordination with dynamic expertise and resource allocation.
- LLM-based Multi-Agent Systems: Early frameworks such as AutoGen and MetaGPT established structured workflows for orchestrating multiple LLM agents.AutoGen emphasized flexible agent interactions, while MetaGPT incorporated software-development principles for collaboration.
- LLM-based Multi-Agent Systems: AgentScope and MegaAgent improve reliability or scalability but continue to use centralized paradigms in which a master agent delegates tasks.These systems therefore retain limitations associated with centralized control.
- Evolutionary Agent Systems: Evolutionary research has separately optimized prompts, interaction topologies, and agent roles or personas.These directions target input design, communication patterns, task allocation, collaboration efficiency, or specialization.
- Evolutionary Agent Systems: Many evolutionary approaches focus on individual adaptation and remain centralized, limiting scalability and dynamic adaptability.Their system-level design processes improve orchestration but do not fully resolve decentralized collective coordination.
- AgentNet: AgentNet integrates evolutionary learning with decentralized control so heterogeneous agents can evolve roles, adapt strategies in real time, and collaborate across large-scale systems.This combination distinguishes it from approaches designed primarily for single-agent adaptation.
3 Methodology
AgentNet combines decentralized routing, dynamic task allocation, adaptive learning, and evolving network connectivity to coordinate specialized agents without a central controller. Agents route, split, or execute tasks using local knowledge, update capabilities and memories from experience, and prune ineffective connections.
- Architecture: AgentNet represents agents and directed communication connections as a graph, with each agent containing a router and executor.The router makes routing decisions, while the executor responds to execution queries; both use memory modules supported by RAG.
- Decentralized Network Topology: Decision-making is decentralized because each agent independently routes tasks using local knowledge and task requirements.This removes reliance on a central controller and distributes control across the network.
- Dynamic Task Allocation: AgentNet allocates tasks dynamically by selecting capable agents and allowing routers to forward, split, or execute tasks.Splitting lets an agent complete matching portions and route remaining subtasks to appropriate agents; execution completes the task without further delegation.
- Dynamic Network Update: The network updates connection weights from collaborative success and prunes edges below threshold θw to improve routing efficiency.The decay factor α balances historical performance with recent interactions, while the success metric evaluates routed tasks.
- Adaptive Learning and Specialization: Agents continuously specialize from task experience without explicit role assignment by updating capabilities and retrieving useful memories.This adaptive learning supports self-organization and adaptation to changing task requirements.
- Dynamic Task Allocation: When splitting tasks, agents forward completed subtask results rather than decomposition reasoning, limiting unnecessary information transfer and error propagation.The design preserves only the outputs needed by downstream agents.
4 Experiment
AgentNet is evaluated across mathematics, logical question answering, API-calling, router ablations, heterogeneity settings, and evolution-phase comparisons. Results indicate competitive or superior performance, with decentralized routing, agent diversity at larger team sizes, and evolution improving outcomes.
- Experimental Setup: The evaluation covers mathematics, logical question answering, and API-calling benchmarks, using custom training and test sets across three task categories.
- Main Results: AgentNet achieves competitive or superior performance across all evaluated tasks and outperforms centralized MetaGPT, which reaches 53.00% accuracy on Logical QA.
- Heterogeneous Agents: With 3 agents, homogeneous teams perform best, whereas with 5 agents, heterogeneous configurations outperform homogeneous teams on BBH.
- Router Effectiveness: AgentNet reaches 82.14% training accuracy and 86.00% testing accuracy on BBH, outperforming randomized routing methods.
- Evolution Ablation: Evolution improves MATH performance from 77.86 to 85.00, function-calling performance from 23.00 to 32.00, and BBH accuracy from 76% to 86%.
- Qualitative Comparison: The ReAct case study attributes an incorrect single-step response to a lack of collective reasoning on a more complex reasoning task.
5 Analysis
Analysis experiments examine scaling, autonomous specialization, and network evolution. Increasing resources yields incremental gains with diminishing returns, while agents develop differentiated abilities and collaboration patterns over time.
- Scaling: Training performance increases from 80.38 with 3 agents and 30 executors to 81.18 with 9 agents and 40 executors.Testing performance fluctuates between 80 and 86 as resources increase.
- Scaling: Increasing agents and executor capacity produces gradual performance improvements, but marginal gains diminish as the system scales.
- Autonomous Specialization: With a fixed executor pool of 40, ability scores vary across reasoning, language, knowledge, and sequence tasks as agent counts change.
- Autonomous Specialization: Specialization becomes clearer from 3 to 9 agents, with particular agents excelling in knowledge and sequence abilities.
- Network Evolution: The five-agent BBH network evolves from uniform fully connected weights of 1.00 to differentiated connection strengths and distinct agent roles.
- Network Evolution: The final evolved network links stronger connections with tighter cooperation and improved task allocation, scalability, and fault tolerance.
6 Limitations and Future Work
AgentNet’s remaining challenges concern performance in heterogeneous agent environments and routing beyond small predefined candidate pools. Future work focuses on adapting coordination and resource allocation while enabling broader agent discovery.
- AgentNet’s performance in heterogeneous agent environments remains an open challenge.Agents may differ in model capabilities, workflows, tools, and available data.
- The effects of agent heterogeneity on task coordination and resource allocation remain insufficiently understood.Addressing these variations efficiently is important for scalability and effectiveness in complex environments.
- Router decision-making for exploration and discovery requires further study, especially as agent pools grow larger and more diverse.Current routing selects agents from a relatively small predefined candidate pool.
- Future routing mechanisms should identify and delegate tasks to suitable agents across large, heterogeneous pools.The proposed direction targets more autonomous and accurate delegation.
- Incentives for exploring beyond predefined candidates could enable discovery of new agents or specialized capabilities.The authors identify this as a promising direction for improving adaptability, scalability, performance, and autonomy.
7 Conclusion
AgentNet addresses limitations of centralized multi-agent systems through decentralized coordination, dynamic task allocation, and adaptive learning. The authors report improved task efficiency, adaptability, and specialization, alongside privacy-preserving cooperation across organizations.
- AgentNet combines decentralized architecture, dynamic task allocation, and adaptive learning to address centralized multi-agent system limitations.The conclusion presents this combination as the framework’s core approach.
- AgentNet improves scalability, fault tolerance, and task efficiency in collaborative environments.
- AgentNet’s privacy-preserving features support secure cooperation across organizations.
- The reported experimental results show improvements in task efficiency, adaptability, and specialization.
- AgentNet is presented as a practical framework for flexible and secure multi-agent systems in dynamic, real-world settings.
Ethics Statement
The study reports no human participants, animal subjects, or personal-data use. Its datasets and benchmarks are publicly available and used under their respective licenses, so no ethics approval was required.
- The study did not involve human participants, animal subjects, or personal data.
- All datasets and benchmarks were publicly available and used according to their respective licenses.
- No ethics approval was required.