Hierarchical Decentralized Multi-Agent Coordination with Privacy-Preserving Knowledge Sharing: Extending AgentNet for Scalable Autonomous Systems

Goutham Nalagatla · 2025

Read Hierarchical Decentralized Multi-Agent Coordination with Privacy-Preserving Knowledge Sharing: Extending AgentNet for Scalable Autonomous Systems by Goutham Nalagatla (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-00614.txt.

Introduction

Multi-agent systems let language-model agents work as a team with no central controller. Goutham Nalagatla extends AgentNet in 2025 with hierarchical coordination and privacy-preserving knowledge sharing, aiming at scalable autonomous systems. AgentNet showed fully decentralized coordination through dynamic directed acyclic graph topologies.

Limitations remain: scalability with large agent populations, message overhead, missing privacy guarantees, and weak use of shared compute. The extension attacks those gaps with hierarchy and protected sharing. As agent groups grow, coordination that preserves privacy becomes a must for use in business and public systems. Hierarchy gives the coordinator-free system a way to organize very large agent groups. Privacy-preserving sharing lets agents cooperate without exposing their raw knowledge. For firms that plan fleets of such aides, the gain is a way to grow. More aides can join, share what they know, and still keep each firm's facts safe.

Catalog record

CREATORGoutham Nalagatla
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE13 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2512-00614

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Hierarchical Decentralized Multi-Agent Coordination with Privacy-Preserving Knowledge Sharing: Extending AgentNet for Scalable Autonomous Systems Goutham Nalagatla Abstract Decentralized multi-agent systems have shown promise in enabling autonomous collaboration among LLM-based agents. While AgentNet demonstrated the feasibility of fully decentralized coordination through dynamic DAG topologies, several limitations remain: scalability challenges with large agent populations, communication overhead, lack of privacy guarantees, and suboptimal resource allocation. We propose AgentNet++ , a hierarchical decentralized framework that extends AgentNet with multi-level agent organization, privacy-preserving knowledge sharing via differential privacy and secure aggregation, adaptive resource management, and theoretical convergence guarantees. Our approach introduces cluster-based hierarchies where agents self-organize into specialized groups, enabling efficient task routing and knowledge distillation while maintaining full decentralization. We provide formal analysis of convergence properties and privacy bounds, and demonstrate through extensive experiments on complex multi-agent tasks that AgentNet++ achieves 23% higher task completion rates, 40% reduction in communication overhead, and maintains strong privacy guarantees compared to AgentNet and other baselines. Our framework scales effectively to 1000+ agents while preserving the emergent intelligence properties of the original AgentNet. 1 Introduction Large Language Model (LLM) based multi-agent systems have emerged as a powerful paradigm for solving complex tasks through collaborative intelligence. Traditional centralized orchestration approaches suffer from single points of failure, scalability bottlenecks, and privacy concerns when agents operate across organizational boundaries. AgentNet [ 1 ] introduced a breakthrough framework for fully decentralized agent coordination, where agents form dynamic directed acyclic graph (DAG) networks that evolve based on task demands. While AgentNet demonstrated significant advantages over centralized systems, several critical limitations remain unaddressed: Scalability: As the number of agents grows, the communication complexity in flat DAG topologies becomes prohibitive. Each agent must maintain connections with multiple peers, leading to quadratic growth in message passing. Privacy: While AgentNet minimizes data sharing, it lacks formal privacy guarantees. Agents may inadvertently expose sensitive information through direct knowledge exchange. Resource Efficiency: The framework does not explicitly model agent capabilities or resource constraints, leading to suboptimal task allocation and potential resource contention. Convergence Guarantees: While empirical results are promising, theoretical analysis of convergence and optimality properties is limited. We propose AgentNet++ , a comprehensive extension that addresses these limitations through: 1. Hierarchical Decentralized Architecture: Multi-level agent organization into clusters based on expertise and task similarity, enabling efficient intra-cluster and inter-cluster coordination. 2. Privacy-Preserving Knowledge Sharing: Integration of differential privacy mechanisms and secure aggregation protocols to enable knowledge distillation without exposing raw data. 3. Adaptive Resource Management: Dynamic capability-aware task routing and resource allocation that optimizes system-wide efficiency. 4. Theoretical Guarantees: Formal analysis of convergence, optimality, and privacy properties with provable bounds. Our contributions include: (1) a novel hierarchical decentralized architecture that maintains full autonomy while improving scalability, (2) privacy-preserving protocols for secure knowledge sharing, (3) adaptive resource management algorithms with theoretical guarantees, (4) comprehensive experimental evaluation demonstrating significant improvements over AgentNet and other baselines, and (5) open-source implementation for reproducibility. 2 Related Work 2.1 Decentralized Multi-Agent Systems Decentralized coordination has been extensively studied in multi-agent systems [ 2 , 3 ] . Recent work has explored LLM-based agents in collaborative settings [ 4 , 5 ] . AgentNet [ 1 ] pioneered fully decentralized LLM agent coordination without central controllers, demonstrating emergent collective intelligence through dynamic DAG topologies. 2.2 Hierarchical Multi-Agent Systems Hierarchical organization has been shown to improve scalability in distributed systems [ 6 , 7 ] . However, existing approaches often require central coordination or predefined hierarchies, limiting adaptability. 2.3 Privacy in Multi-Agent Learning Differential privacy [ 8 ] and secure aggregation [ 9 ] have been applied to federated learning, but their integration with decentralized agent coordination remains underexplored. 3 Methodology 3.1 System Architecture AgentNet++ extends AgentNet’s dynamic DAG structure with a hierarchical organization. The system consists of three levels: Level 1 - Individual Agents: Each agent a_{i} maintains: • Local state s_{i} and capability profile c_{i} • Retrieval-based memory M_{i} for task history • Connection set N_{i} of neighboring agents • Privacy budget \epsilon_{i} for differential privacy Level 2 - Agent Clusters: Agents self-organize into clusters C_{k} based on: • Task similarity: agents working on related tasks • Expertise complementarity: agents with complementary skills • Communication efficiency: agents with low-latency connections Each cluster C_{k} has a dynamic cluster head h_{k} selected through decentralized consensus. Level 3 - Inter-Cluster Coordination: Clusters coordinate through their cluster heads, forming a meta-graph G_{meta} where nodes are clusters and edges represent inter-cluster communication. 3.2 Hierarchical Task Decomposition and Routing Tasks arrive at the system and are decomposed hierarchically: Algorithm 1 Hierarchical Task Routing 0: Task T , Agent set A , Cluster set C 0: Assigned agent a^{*} 1: Decompose T into subtasks \{T_{1},\ldots,T_{n}\} 2: for each subtask T_{i} do 3: Identify candidate clusters C_{cand}=\{C_{k}:\text{expertise}(C_{k})\cap\text{requirements}(T_{i})\neq\emptyset\} 4: Select cluster C^{*}=\arg\max_{C_{k}\in C_{cand}}\text{score}(C_{k},T_{i}) 5: Within C^{*} , select agent a^{*}=\arg\max_{a_{j}\in C^{*}}\text{capability}(a_{j},T_{i}) 6: Assign T_{i} to a^{*} 7: end for The scoring function considers: \text{score}(C_{k},T_{i})=\alpha\cdot\text{expertise\_match}(C_{k},T_{i})+\beta\cdot\text{resource\_availability}(C_{k})-\gamma\cdot\text{load}(C_{k}) (1) where \alpha , \beta , \gamma are weighting parameters. 3.3 Privacy-Preserving Knowledge Sharing To enable knowledge sharing while preserving privacy, we employ two mechanisms: Differential Privacy: When agent a_{i} shares knowledge K_{i} with cluster members, it applies (\epsilon,\delta) -differential privacy: K_{i}^{priv}=K_{i}+\mathcal{N}(0,\sigma^{2}\cdot\Delta K_{i}) (2) where \Delta K_{i} is the sensitivity of knowledge K_{i} and \sigma^{2}=\frac{2\ln(1.25/\delta)}{\epsilon^{2}} . Secure Aggregation: For aggregating knowledge from multiple agents, we use a secure aggregation protocol: K_{agg}=\sum_{a_{i}\in C_{k}}w_{i}\cdot K_{i}^{priv}\pmod{p} (3) where w_{i} are weights and p is a large prime for modular arithmetic. 3.4 Adaptive Resource Management Each agent maintains a capability vector c_{i}\in\mathbb{R}^{d} encoding: • Computational resources: CPU, memory, GPU availability • Domain expertise: proficiency in different task domains • Communication bandwidth: network capacity The resource manager dynamically updates capabilities based on: c_{i}^{t+1}=c_{i}^{t}+\eta\cdot\nabla_{c_{i}}\mathcal{L}_{task}(a_{i},T) (4) where \mathcal{L}_{task} is the task loss and \eta is the learning rate. 3.5 Cluster Formation and Evolution Clusters form and evolve through a decentralized process: Algorithm 2 Decentralized Cluster Formation 0: Agent set A , Similarity threshold \theta 0: Cluster set C 1: Initialize: each agent a_{i} starts as singleton cluster 2: repeat 3: for each agent a_{i} do 4: Compute similarity to all clusters: sim(a_{i},C_{k}) 5: if \max_{k}sim(a_{i},C_{k})>\theta then 6: Join cluster C^{*}=\arg\max_{k}sim(a_{i},C_{k}) 7: else 8: Form new cluster or remain independent 9: end if 10: end for 11: Update cluster heads through consensus 12: until convergence Similarity is computed as: sim(a_{i},C_{k})=\lambda_{1}\cdot\text{task\_similarity}(a_{i},C_{k})+\lambda_{2}\cdot\text{expertise\_complementarity}(a_{i},C_{k})-\lambda_{3}\cdot\text{communication\_cost}(a_{i},C_{k}) (5) 4 Theoretical Analysis 4.1 Convergence Guarantees Theorem 1 (Task Completion Convergence): Under Assumptions A1-A3 (bounded task complexity, finite agent capabilities, connected communication graph), the hierarchical task routing algorithm converges to a task assignment with probability 1, and the expected completion time is bounded by O(\log|A|\cdot\log|T|) where |A| is the number of agents and |T| is the number of tasks. Proof Sketch: The hierarchical structure reduces the search space from O(|A|^{|T|}) to O(|C|^{|T|}\cdot\max_{k}|C_{k}|^{|T|}) where |C| is the number of clusters. The cluster-based routing ensures that each task is assigned to a capable agent within finite iterations. The logarithmic factors arise from the hierarchical search and consensus mechanisms. 4.2 Privacy Guarantees Theorem 2 (Differential Privacy): The knowledge sharing mechanism satisfies (\epsilon,\delta) -differential privacy where \epsilon=\sum_{i=1}^{n}\epsilon_{i} for n sharing events, and \delta=\sum_{i=1}^{n}\delta_{i} . Proof: Follows from composition theorem of differential privacy [ 10 ] . Each knowledge sharing event adds noise calibrated to \epsilon_{i} , and the total privacy

Read the complete text