Dynamic LLM Agent Network (DyLAN)

• Dynamic LLM-Powered Agent Network (DyLAN) is a multi-agent architecture that orchestrates specialized LLM agents with adaptive, task-driven topologies.
• It integrates dynamic agent selection, structured communication protocols, and external tools like graph neural networks to enhance performance in domains such as materials discovery and wireless networks.
• Empirical evaluations demonstrate significant improvements in speed, utilization, and task completion rates compared to traditional static or single-agent systems.

Dynamic LLM-Powered Agent Network (DyLAN) constitutes a class of collaborative multi-agent architectures wherein LLM agents, often augmented by external reasoning modules (e.g., graph neural networks), assemble into adaptively configured teams or topologies to solve tasks across various domains. DyLAN frameworks are defined by their dynamic agent selection, communication patterns, and orchestration mechanisms, enabling both structural and functional adaptiveness, task-driven optimization, and scalable deployment. Recent realizations span automated materials discovery, task-oriented collaboration, wireless resource management, cross-domain networking, and document understanding, demonstrating superior performance and efficiency over static multi-agent or single-agent baselines.

1. Architectural Principles and System Overview

DyLAN implementations operate by instantiating a set of LLM-driven agents—each assigned specialized roles (e.g., planning, reviewing, coding, analysis)—and orchestrating their interactions in a dynamic, event-driven manner (Ghafarollahi et al., 2024, Liu et al., 2023). Typically, a central coordinator ("Assistant") agent decomposes user queries into subtasks, dispatches those to appropriate agents or external tools, and iteratively integrates feedback to refine and synthesize final outputs.

Agents communicate via structured protocols (e.g., JSON-based message passing or graph-based workflows), enabling both sequential and parallel execution, with runtime adaptiveness to task demands and intermediate results (Ghafarollahi et al., 2024, Xia et al., 22 Jul 2025). Agent pools are often complemented by distributed memory structures for storing global state, task history, and intermediate outputs (Tong et al., 2 May 2025).

Table: Canonical DyLAN Agent Roles and Coordination

Role	Function	Example Implementation
Assistant	Orchestration, dispatching	"CALL_TOOL"/integration agent
Planner	Plan synthesis and decomposition	Step-by-step task planner
Reviewer	Plan or output refinement	Critique/approve plans
Coder	Code generation/exec	Python plot generator
Analyzer	Modal analysis, insight summary	Multi-modal output reviewer
External Tool	Physics or domain inference	GNN, convex solver

DyLAN instantiations vary by domain (materials science, wireless networks, document QA, etc.) but consistently exhibit modularity, dynamic role assignment, and explicit agent interaction topologies.

2. Dynamic Topology and Team Optimization

The hallmark of DyLAN is its dynamic (often task-adaptive) topology. Unlike fixed chain or fully connected graphs, DyLAN selects its agent team and configures inter-agent communication structure per task, sample, or runtime metric (Liu et al., 2023, Jiang et al., 9 Oct 2025, Leong et al., 31 Jul 2025). Team optimization is accomplished via agent selection algorithms, commonly employing importance scores based on peer ratings during a calibration phase:

• Agent Importance Score: In unsupervised multi-round trials, each agent’s contribution is quantified via backward aggregation of peer ratings. The top-k agents are chosen for the subsequent task-solving phase (Liu et al., 2023).
• Iterative Pruning and Consensus: During inference, agents may be dynamically pruned by rankers evaluating intermediate outputs, while early-stopping is enacted upon consensus (e.g., majority, BLEU threshold) (Liu et al., 2023).

Advanced DyLAN variants leverage generative graph frameworks (such as Guided Topology Diffusion, GTD) that evolve the communication topology through conditional diffusion steered by proxy reward predictors, balancing utility (accuracy), cost (token usage), sparsity, and robustness (Jiang et al., 9 Oct 2025). Alternatively, reinforcement learning agents (e.g., A2C in DynaSwarm) optimize edge parameters in the agent graph, while parameter-efficient LLM selectors score and pick the optimal structure for each input (Leong et al., 31 Jul 2025).

3. Coordination Protocols and Orchestration Mechanisms

DyLAN systems utilize event-driven, modular orchestration protocols. The central coordinator agent (e.g., Assistant or Orchestrator) interacts with specialized agents and external tools through standardized calls, responses, and iterative feedback loops (Ghafarollahi et al., 2024, Xia et al., 22 Jul 2025). This design supports:

• Task Decomposition: Queries are mapped to multi-step plans (generate, predict, plot, analyze), inspected by Reviewer agents, then enacted via further dispatch (Ghafarollahi et al., 2024).
• Adaptive Routing: Dynamic task router modules allocate subtasks based on agent confidence/workload (“softmax over confidence minus workload”), supporting feasible load balancing and self-healing (Xia et al., 22 Jul 2025).
• Bidirectional Critique: Feedback buses (pub/sub channels) carry structured critiques enabling upstream correction and agent reassignment (Xia et al., 22 Jul 2025).
• Parallelism and Competition: Ambiguous subtasks invoke parallel agents and evaluator-driven output selection to maximize diversity and factual coverage (Xia et al., 22 Jul 2025).

Agent-to-agent and agent-to-tool communication is encoded formally (e.g., JSON payloads, Chain-of-Identity prompt templates), with shared memory or state graphs facilitating global context and coordinated decision-making (Ghafarollahi et al., 2024, Tong et al., 2 May 2025).

4. Integration of External Reasoning Modules

Robust DyLAN platforms frequently incorporate domain-specific external reasoning modules—most prominently GNN-powered property predictors, convex optimization tools, or digital twins (Ghafarollahi et al., 2024, Tong et al., 2 May 2025, Zhang et al., 1 Apr 2025):

• Materials Discovery: A GNN physics tool predicts atomic-scale properties (Peierls barrier, interaction energy) in seconds, providing >34,000× speedup over ab initio simulations; Python coder agents plot ternary diagrams, with multi-modal analyzers summarizing trends (Ghafarollahi et al., 2024).
• Wireless Networks: Agents (e.g., slice allocator, bandwidth allocator) interact via a graph-based runtime (LangGraph), using chain-of-thought and retrieval-augmented prompts. External tools verify or refine allocations, with local optimization routines (e.g., bisection) employed for QoS satisfaction (Tong et al., 2 May 2025).
• Optical Network Autonomy: Agents dynamically spawn/retire per failure event or training epoch. Planner agents orchestrate subtasks, while task agents use retrieval-augmented models for probing physical-layer status, achieving 98% task completion (Zhang et al., 1 Apr 2025).

Such integrations enable DyLAN to transcend LLM-only reasoning constraints by fusing statistical, symbolic, and physics-based inference.

5. Algorithmic Formulations and Performance Metrics

DyLAN frameworks define explicit scoring and utility functions for agent output ranking and subtask prioritization. Example mathematical formulations include:

• Utility Function (materials): $$U(\mathrm{comp}) = w_b\,\hat{y}_\mathrm{barrier} + w_e\,\hat{y}_{\Delta U} - \lambda\,\mathrm{cost}$$, with weights for different output properties and (often negligible) computation cost (Ghafarollahi et al., 2024).
• Wireless Throughput: Individual user rates $$\Gamma_n(B_n) = \alpha B_n \log_2(1 + 10^{\eta_n/10})$$, with total throughput and allocation subject to QoS and feasibility constraints (Tong et al., 2 May 2025).
• Proxy Reward Model (topology design): A GNN surrogate $\mathcal{P}_\phi(G, c) = [\hat u, \hat c]$ predicts utility/cost of candidate topologies, steering the diffusion sampling loop (Jiang et al., 9 Oct 2025).

Empirical evaluations consistently show DyLAN architectures outperform static baselines—whether single-agent CoT, naive multi-agent, or fixed-template MAS:

Table: Representative DyLAN Performance Gains

Domain	Baseline	DyLAN Variant	Relative Gain
Alloy Discovery	NEB brute-force	GNN-driven multi-agent	>34,000× speed-up
Wireless Slicing	Prompt-based	WirelessAgent/LangGraph	+44.4% utilization
Document Understanding	Static multi-agent	Adaptive+Parallel DyLAN	+29% factual cover.
Optical Network Automation	Single-Agent LLM	AutoLight DyLAN	3.2× completion rate

Metrics include computational cost (number of API calls, GPU/hours), mean absolute error (MAE), R² accuracy, factual coverage, redundancy penalty, and latency (Ghafarollahi et al., 2024, Tong et al., 2 May 2025, Xia et al., 22 Jul 2025, Zhang et al., 1 Apr 2025).

6. Limitations, Trade-offs, and Future Directions

DyLAN architectures face several operational constraints:

• Overhead: Messaging and agent spawning can increase latency and resource use, especially in high-fanout or real-time control settings; LLM invocation latency may challenge sub-second loops (Xia et al., 22 Jul 2025, Zhang et al., 1 Apr 2025).
• Agent Scoring Bias: Dependence on agent importance scores or peer ratings may be brittle in low-data or adversarial regimes (Liu et al., 2023).
• Static Parameters: Fixed scoring weights (α, β, γ) may not generalize, motivating meta-learning or adaptive tuning (Xia et al., 22 Jul 2025).
• Consistency and Drift: Shared memory and retrieval corpora require regular updating to prevent stale state, particularly as hardware or data evolves (Zhang et al., 1 Apr 2025).

Future research avenues include reinforcement-learning-based planner tuning, hierarchical dynamic graph sampling (meta-RL), human-in-the-loop auditing for sensitive domains, multi-modal and embodied agent integration, and federated deployments with privacy-preserving protocols (Leong et al., 31 Jul 2025, Zhang et al., 1 Apr 2025).

7. Domain-Specific Instantiations and Generalization

DyLAN’s versatility is evidenced by its deployment across diverse technical domains:

• Materials Science: Automated exploration and optimization of metallic alloys using LLM/GNN synergy, massive acceleration over direct simulations (Ghafarollahi et al., 2024).
• Wireless Networks: Autonomous resource management via cognitive agent modules, near-optimal throughput with substantial utilization gains (Tong et al., 2 May 2025).
• Optical Networking: Level-4 autonomy with field-trial validation; integrated cross-domain, cross-layer agents supporting distributed AI training workflows (Zhang et al., 1 Apr 2025).
• Document Understanding: Adaptive, parallel multi-agent systems yielding improved factual and compliance metrics over static frameworks (Xia et al., 22 Jul 2025).

These applications validate the theoretical and empirical advantages of dynamic agent selection, adaptive topology, and modular orchestration intrinsic to DyLAN. Generalizable methodological themes include event-driven agent coordination, task-aware graph design, and reward-guided multi-agent optimization.

---

DyLAN frameworks establish a rigorous foundation for collaborative, scalable, and efficient multi-LLM agent systems, unifying adaptive orchestration, principled communication design, and external reasoning integration. Continued research focuses on enhancing robustness, cost-efficiency, and application breadth across increasingly complex real-world tasks.