AgentNet++: Scalable Multi-Agent Framework

• The paper introduces a hierarchical decentralized framework that reduces communication overhead via multi-level clustering and optimized task routing.
• It employs differential privacy and secure aggregation to enable privacy-preserving knowledge sharing, ensuring formal theoretical guarantees and reduced complexity.
• Adaptive resource management integrated with semantic-aware coordination (as seen in SANNet) enhances task success and efficient multi-agent collaboration at scale.

AgentNet++ is a hierarchical decentralized framework for multi-agent coordination, designed to resolve the scalability, privacy, and resource management limitations inherent to the original AgentNet. This architecture enables autonomous collaboration among large populations of LLM-based agents, introducing structured multi-level clustering, privacy-preserving knowledge sharing via differential privacy and secure aggregation, adaptive resource management, and formal theoretical guarantees. AgentNet++ maintains full decentralization while supporting efficient multi-agent task routing and emergent intelligence properties at scale (Nalagatla, 29 Nov 2025). The framework also subsumes semantic-aware agentic AI networking systems such as SANNet by accommodating cross-layer semantic goal discovery and conflict-resolved orchestration within a unified hierarchical model (Xiao et al., 25 May 2025).

1. Architectural Principles and Motivation

AgentNet++ is motivated by bottlenecks in flat DAG topologies used by AgentNet, where communication complexity scales quadratically ($O(N^2)$) with the number of agents, preventing scalability to large agent populations. Privacy concerns arise from the lack of guarantees in peer-to-peer knowledge exchange, alongside suboptimal resource allocation due to the absence of explicit modeling for agent capabilities and resource constraints. AgentNet++ introduces:

• Hierarchical decentralized organization: Agents self-organize into clusters, forming a three-level hierarchy (individual agent, intra-cluster, inter-cluster).
• Differential privacy and secure aggregation for knowledge distillation.
• Adaptive, dynamic resource management and capability-aware task routing.
• Formal theoretical analysis of convergence, privacy, and communication complexity.

This design enables scalable coordination, reduced communication overhead, and privacy guarantees, while preserving decentralization.

2. Hierarchical Multi-Agent Organization

Formally, agents $A = \{a_1, \ldots a_N\}$ are partitioned into $K$ clusters $C = \{C_1, \ldots, C_K\}$, each with a dynamic cluster head $h_k \in C_k$, via a decentralized clustering mapping $\varphi: A \to C$. This organization yields:

• Level 1 (Individual): Agents maintain local state $s_i$, capability profile $c_i \in \mathbb{R}^d$, private memory $M_i$, neighbor set $N_i$, and privacy budget $(\epsilon_i, \delta_i)$.
• Level 2 (Clusters): Sets $C_k$ with elected head $h_k$.
• Level 3 (Inter-Cluster): Meta-graph $G_{\text{meta}}$ among cluster heads, dictating higher-level communication.

Cluster formation is iterative and similarity-driven. Each agent computes a similarity metric:

$$\text{sim}(a_i, C_k) = \lambda_1\cdot \text{task\_similarity}(a_i, C_k) + \lambda_2\cdot \text{expertise\_complementarity}(a_i, C_k) - \lambda_3\cdot \text{communication\_cost}(a_i, C_k)$$

Agents join clusters if the similarity exceeds a threshold $\theta$; otherwise, new clusters are formed. Cluster heads are chosen by decentralized consensus, e.g., via max degree.

Hierarchical task routing involves decomposing tasks into subtasks, scoring clusters by expertise match, resource availability, and load:

$$\text{score}(C_k, T_i) = \alpha\cdot \text{expertise\_match}(C_k, T_i) + \beta\cdot \text{resource\_availability}(C_k) - \gamma\cdot \text{load}(C_k)$$

The highest-scoring cluster and then the most capable agent within that cluster are assigned the subtask. Knowledge distillation occurs via periodic sharing and aggregation of privatized agent summaries.

3. Privacy-Preserving Knowledge Sharing

Privacy mechanisms in AgentNet++ rely on differential privacy and secure aggregation. Each agent privatizes its knowledge vector $K_i$ by adding Gaussian noise:

$$K_i^{\text{priv}} = K_i + \mathcal{N}(0, \sigma^2 I), \quad \sigma^2 = \frac{2\ln(1.25/\delta_i)}{\epsilon_i^2}$$

This enforces $(\epsilon_i, \delta_i)$-differential privacy per agent per sharing event. Secure aggregation at the cluster level, via threshold secret sharing, ensures only the aggregate cluster summary $K_{\text{agg}}$ can be reconstructed, preventing leakage of individual agents' private summaries ($K_i^{\text{priv}}$):

$$K_{\text{agg}} \equiv \sum_{a_i \in C_k} w_i K_i^{\text{priv}} \mod p$$

Privacy budgets compose additively across $n$ rounds:

$$\epsilon_{\text{total}} = n \epsilon_0, \quad \delta_{\text{total}} = n \delta_0$$

Secure aggregation is a post-processing step, which does not impact differential privacy guarantees (cf. Dwork & Roth 2014).

4. Adaptive Resource Management

Each agent maintains a capability vector $$c_i = [\text{CPU}_i, \text{RAM}_i, \text{GPU}_i, \text{domain}_1^i, \ldots, \text{domain}_m^i, \text{bandwidth}_i]^T \in \mathbb{R}^d$$. The global optimization goal is:

$$\min_{\text{assignments}} \;\; \mathcal{L}_{\text{total}} = \sum_{T \in S}\mathcal{L}_{\text{task}}(a_T, T) + \lambda \cdot \text{CommunicationCost}$$

Upon subtask completion, agents update their profiles:

$$c_i^{t+1} = c_i^t + \eta \cdot \nabla_{c_i}\mathcal{L}_{\text{task}}(a_i,T)$$

This gradient-based update balancing domain proficiency and resource usage enables dynamic reallocation of resources and self-optimization, supporting time-varying demand scenarios and heterogeneous agent capabilities.

5. Theoretical Guarantees and Complexity Analysis

AgentNet++ provides formal convergence, privacy, and communication complexity guarantees. Under bounded task complexity, finite agent capabilities, and a connected communication graph:

• Task Assignment Convergence: Hierarchical routing converges in expected time $O(\log |C| \cdot \log |T|)$.
• Differential Privacy Composition: After $n$ knowledge-sharing rounds, the framework satisfies $(\epsilon=\sum \epsilon_i, \delta=\sum \delta_i)$-DP.
• Communication Complexity: Balanced clustering (with $|C| \approx |C_k| \approx \sqrt{N}$) yields total communication $O(N^{1.5})$, compared to $O(N^2)$ for flat AgentNet topologies.

The reduction of branching factor and communication cost is a direct consequence of hierarchical organization and balanced partitioning across clusters.

6. Empirical Evaluation and SANNet Integration

Experimental results on benchmarks covering complex reasoning, distributed information gathering, and dynamic task allocation demonstrate tangible advantages:

• Task Completion Rate: AgentNet++ achieves 87.3%, outperforming AgentNet (71.0%) and centralized orchestrators (60.2%).
• Communication Overhead: 40% reduction compared to AgentNet at $N=500$, with overhead scaling as $O(N^{1.5})$ vs. $O(N^2)$.
• Scalability: >85% task success for $N>1000$ agents; AgentNet degrades past $N=200$.
• Privacy-Utility Tradeoff: With $(\epsilon=1.0,\delta=10^{-5})$, only 2.1% drop in accuracy.

As an AgentNet++ system, SANNet introduces semantic goal inference, cross-layer orchestration, and dynamic conflict resolution for mobile networking platforms (Xiao et al., 25 May 2025). The semantic goal inference module maps user inputs $x \in X$ to discrete semantic goals $g \in G$, decomposed into cross-layer requirements. SANNet’s Agent Controller orchestrates application, network, and physical layer agents by dynamically adjusting Pareto weights through a stochastic update mechanism (Algorithm 1), optimizing conflicting objectives toward Pareto-stationary equilibria.

In a 5G RAN+5GS core prototype, SANNet demonstrated up to 63% reduction in multi-agent conflict error (C-error), 31% improvement in video QoE, and a 16 percentage-point increase in goal-achievement rate over static multi-agent baselines. Theoretical guarantees include bounded C-error and G-error with prescribed convergence rates under standard Lipschitz and stability assumptions.

7. Significance, Implications, and Future Directions

AgentNet++ generalizes decentralized agentic frameworks to support scalable, privacy-guaranteed, and resource-adaptive coordination across heterogeneous agent populations. The hierarchical architecture and formal treatment of privacy and convergence underpin its ability to scale to over 1000 agents while maintaining high task success and communication efficiency. The integration of semantic-aware modules exemplified by SANNet expands AgentNet++ into cross-layer domains such as autonomous networking.

A plausible implication is that future research can extend AgentNet++'s hierarchical and privacy-preserving coordination protocols to additional large-scale multi-agent environments, such as internet-of-things, automated supply chains, or smart infrastructure, leveraging both theoretical guarantees and practical performance demonstrated in empirical studies (Nalagatla, 29 Nov 2025, Xiao et al., 25 May 2025).