PaperlayerView exact revision on arXiv

Source-linked AI summary

Toward Edge General Intelligence with Agentic AI and Agentification: Concepts, Technologies, and Future Directions

Ruichen Zhang, Guangyuan Liu, Yinqiu Liu, Changyuan Zhao, Jiacheng Wang, Yunting Xu, Dusit Niyato, Jiawen Kang, Yonghui Li, Shiwen Mao, Sumei Sun, Xuemin Shen, Dong In Kim

arXiv:2508.18725v1cs.NIcs.IT

TL;DR

Traditional edge intelligence is limited by static models, centralized dependence, and weak adaptation in dynamic, resource-constrained environments. The paper surveys Agentic AI and agentification for edge general intelligence, covering their perception–reasoning–action architecture, enabling design pillars, and representative deployments. Across the reported case studies, Agentic AI methods show improved energy consumption, intent fulfillment, accumulated return, and human-centric QoE, while the paper identifies cloud reliance and limited adaptability as important boundaries of traditional approaches.

  • Problem

    Traditional edge intelligence uses static, task-specific models and centralized control, limiting adaptation and robustness in dynamic, heterogeneous, and resource-constrained edge environments.

  • Method

    The paper presents a comprehensive survey of Agentic AI and agentification for edge general intelligence, organizing architectures, enabling technologies, design pillars, and deployment case studies.

  • Results

    The reported case studies show improvements including up to 6.4% lower final energy consumption, 14.8% higher task success, approximately 61% higher accumulated return, and 27.3% higher human-centric QoE.

  • Takeaways & Limitations

    Agentic AI provides an edge-oriented framework for autonomous perception, reasoning, planning, action, collaboration, and continual adaptation across diverse edge applications.

  • Takeaways & Limitations

    Traditional edge approaches remain constrained by heavy cloud reliance and limited adaptability to changing network conditions or user behaviors.

Abstract

from arXiv · show

The rapid expansion of sixth-generation (6G) wireless networks and the Internet of Things (IoT) has catalyzed the evolution from centralized cloud intelligence towards decentralized edge general intelligence. However, traditional edge intelligence methods, characterized by static models and limited cognitive autonomy, fail to address the dynamic, heterogeneous, and resource-constrained scenarios inherent to emerging edge networks. Agentic artificial intelligence (Agentic AI) emerges as a transformative solution, enabling edge systems to autonomously perceive multimodal environments, reason contextually, and adapt proactively through continuous perception-reasoning-action loops. In this context, the agentification of edge intelligence serves as a key paradigm shift, where distributed entities evolve into autonomous agents capable of collaboration and continual adaptation. This paper presents a comprehensive survey dedicated to Agentic AI and agentification frameworks tailored explicitly for edge general intelligence. First, we systematically introduce foundational concepts and clarify distinctions from traditional edge intelligence paradigms. Second, we analyze important enabling technologies, including compact model compression, energy-aware computing strategies, robust connectivity frameworks, and advanced knowledge representation and reasoning mechanisms. Third, we provide representative case studies demonstrating Agentic AI's capabilities in low-altitude economy networks, intent-driven networking, vehicular networks, and human-centric service provisioning, supported by numerical evaluations. Furthermore, we identify current research challenges, review emerging open-source platforms, and highlight promising future research directions to guide robust, scalable, and trustworthy Agentic AI deployments for next-generation edge environments.

I. INTRODUCTION

The paper frames Agentic AI and agentification as a shift from static, task-specific edge intelligence toward autonomous systems that perceive, reason, plan, act, and adapt across heterogeneous edge environments. It surveys an edge-oriented methodology built around four design pillars and illustrates the approach through diverse applications and case studies.

  • Motivation: Traditional edge intelligence relies on static, task-specific models that struggle with mission changes, environmental disturbances, and dynamic multimodal edge conditions.These limitations motivate edge general intelligence, which emphasizes broader cognitive capabilities and continual adaptation.
  • Agentic AI paradigm: Agentic AI operates through continuous perception–reasoning–action loops, enabling context interpretation, deliberative planning, goal-driven decisions, and proactive adaptation.Its architecture includes Perception, Memory, Reasoning, and Action modules connected by memory management and retrieval, with execution on-device or offloaded.
  • Research gap: Existing studies address isolated components or applications, but a unified, edge-oriented methodology connecting multimodal agents to deployment sites and networking remains limited.The survey positions its framework as a systematic treatment of edge general intelligence rather than a collection of disconnected use cases.
  • Contributions: The paper contributes a comprehensive edge-focused survey, use-case analyses, deployment guidance, and discussion of open-source frameworks, challenges, and future directions.Applications include cooperative UAV swarms, adaptive vehicular networks, intent-driven networking, and edge robotics or human-centric services.
  • Design pillars: The survey organizes practical Agentic AI design around compactness, efficiency, migration, and knowledge and reasoning.Representative mechanisms include LoRA, distillation, quantization, early exits, offloading, structured memory, retrieval-augmented generation, and tool use.

C. Paper Organization

The paper organizes its survey around the foundations, enabling technologies, applications, challenges, frameworks, and future directions of edge general intelligence. It contrasts traditional static, cloud-reliant edge intelligence with more adaptive and autonomous approaches.

  • The survey proceeds from Agentic AI foundations to enabling technologies, application scenarios, system challenges, open-source frameworks, and future research directions.
  • Traditional edge intelligence relies on static, task-specific models that struggle with dynamic, multimodal environments and changing mission requirements.
  • Cloud dependence in conventional edge systems introduces latency, bandwidth bottlenecks, and single points of failure, limiting scalability and robustness.
  • Traditional models also lack continuous adaptation and often neglect tight memory, compute, and power constraints, causing energy and responsiveness inefficiencies.
  • The paper frames the transition from passive IoT devices toward an Internet of Agents whose edge nodes perceive, reason, plan, and act independently in real time.

C. Agentic AI

Agentic AI evolves from reactive rule-based systems through task-specific DRL and reasoning-capable LLM agents toward autonomous systems with memory, planning, tool use, and collaboration. Its defining principle is goal-driven autonomy through continuous perception–reasoning–action loops.

  • Definition of Agentic AI: Agentic AI systems operate through continuous perception–reasoning–action loops, enabling goal-driven autonomy and proactive task execution.
  • Definition of Agentic AI: Localized inference and planning, memory-driven continuous learning, and model compression address cloud dependence, adaptation, and computational-efficiency limitations.
  • Key Developments of Agentic AI: The evolution progresses from rule-based agents to DRL-driven agents, LLM-driven agents, and ultimately fully autonomous Agentic AI systems.
  • Key Developments of Agentic AI: Rule-based agents are reactive and effective mainly in static, narrowly defined scenarios, while DRL agents improve adaptability but remain task-specific and lack explicit reasoning.
  • Key Developments of Agentic AI: LLM-driven agents support general reasoning, multi-step planning, and structured tool interactions through frameworks such as Codex, ReAct, and Toolformer.
  • Definition of Agentic AI: Agentic systems integrate autonomy, contextual memory, explicit reasoning, long-term planning, proactive decisions, modular collaboration, and dynamic feedback.

D. Comprehensive Framework of Agentic AI

The comprehensive framework organizes Agentic AI into interconnected perception, memory, reasoning, and action components within an iterative workflow. Foundation models, retrieval, planning, tools, and environmental feedback support autonomous adaptation in edge deployments.

  • Architecture: The architecture comprises Perception, Memory, Reasoning, and Action modules connected by explicit memory management and retrieval, with execution on-device or offloaded to edge or cloud.
  • Perception: Perception integrates textual, visual, and auditory inputs, while multimodal sensor fusion can support context-aware interpretation of complex environments.
  • Comprehension: LLMs provide semantic comprehension and reasoning that interpret complex instructions, decompose high-level tasks, and generate structured plans.
  • Action: External tools and APIs extend agents' capabilities by providing computational resources, databases, and specialized knowledge during reasoning and action.
  • Memory: Memory and retrieval mechanisms such as RAG retain historical knowledge and dynamically retrieve contextually relevant information from vectorized knowledge bases.
  • Reasoning: Explicit planning and reasoning, including CoT and symbolic techniques, enable long-horizon strategy assessment and systematic decomposition of complex goals.
  • Iterative Workflow: The iterative workflow captures multimodal data, builds semantic context, self-plans, refines actions through reasoning and memory, executes tools, and feeds outcomes back for self-refinement.
  • Architecture: Integrating these components establishes a cognitive architecture intended to adapt autonomously to dynamic edge environments and advance edge general intelligence.

III. KEY ENABLERS OF AGENTIC AI FOR EDGE GENERAL INTELLIGENCE

Agentic AI for edge general intelligence depends on coordinated advances in compact models, energy-aware execution, connectivity, collaboration, and reasoning. These enablers support autonomous operation under constrained edge resources and form a reinforcing cycle.

  • Model Compressions and Compact Models: Compact model techniques address edge devices’ stringent memory and computational constraints while preserving Agentic AI’s reasoning and adaptability.Pruning, quantization, low-rank factorization, and knowledge distillation support compact deployment of foundation models and LLMs.
  • Model Compressions and Compact Models: 85% and 85.5% accuracies were achieved by distilled Llama2 (7B) and CodeLlama models, respectively, on the SVAMP arithmetic reasoning dataset.The results surpassed GPT-3.5-turbo and transferred multi-step reasoning into smaller models.
  • Model Compressions and Compact Models: INT4-g128 AWQ quantization delivered lossless performance across 11 vision-language benchmarks.The method reduces model parameters and activations to lower bit-precisions for efficient edge execution.
  • Model Compressions and Compact Models: Up to 20% parameter reduction preserved over 94% of original model performance using LLM-Pruner.The structured pruning method used 50,000 training samples and three hours of fine-tuning.
  • Energy-Efficient Computing: Adaptive early-exit schemes reduced latency by up to 24.6% and energy consumption by 46.5% without accuracy loss.Computation was adaptively distributed between devices and edge servers according to input complexity and environmental constraints.
  • Energy-Efficient Computing: Hardware-aware optimization reduced agent invocation cost by 55.1% and overall energy consumption by up to 23.3%.The approach combines accelerator-aware co-design, pruning, dynamic voltage and frequency scaling, and reinforcement-learning-based scheduling.

C. Connectivity and Collaboration

Connectivity and collaboration enable distributed edge agents to share intelligence, coordinate tasks, and execute complex workloads despite intermittent links and resource constraints. Knowledge representation and reasoning further support anticipation, memory, retrieval, and continual adaptation.

  • Connectivity and Collaboration: Robust connectivity and collaboration are foundational for decentralized agents to cooperate, share intelligence, and execute complex edge tasks.The section emphasizes communication protocols, collaborative decision-making, and adaptive coordination for scalable deployment.
  • Connectivity and Collaboration: Gossip-based algorithms, federated learning protocols, and sparse message passing propagate information with low redundancy under bandwidth constraints.These protocols target reliable coordination in intermittently connected edge scenarios.
  • Connectivity and Collaboration: Multi-agent reinforcement learning, distributed constraint optimization, and graph neural networks support decentralized task allocation and resource coordination.These methods also support planning, reasoning, and efficient invocation of external tools and APIs.
  • Connectivity and Collaboration: Emergent communication allows agents to learn concise, adaptive shared languages that encode task-relevant semantics with minimal communication overhead.The resulting signaling mechanisms strengthen collective collaboration at the edge.
  • Knowledge Representation and Reasoning: RAG, on-device knowledge bases, long-term memory, causal reasoning, world models, and continual learning provide the knowledge and adaptation mechanisms required by edge agents.These techniques support retrieval, historical context, environmental prediction, look-ahead planning, and incremental knowledge updates.
  • Integrated Edge Deployment: Integrated deployment combines model compression, energy-aware execution, hardware optimization, robust communication, decentralized coordination, and reasoning mechanisms.The combined design targets operation within strict edge memory, compute, power, and thermal limits.

A. Agent Frameworks and Platforms

Open-source Agentic AI projects span frameworks, autonomous applications, and domain-specific agents. These systems provide role-based collaboration, visual workflow design, conversational orchestration, autonomous coding, research automation, and benchmarked evaluation.

  • Agent Frameworks and Platforms: Agent frameworks and platforms lower barriers to deploying autonomous agents capable of reasoning, decision-making, and collaboration.The surveyed projects are organized into agent frameworks and platforms, autonomous applications, and domain-specific agents.
  • Agent Frameworks and Platforms: MetaGPT uses role-based multi-agent collaboration for code generation, peer review, iterative refinement, and coordinated execution.Its natural-language programming and task-automation design targets complex software engineering workflows.
  • Agent Frameworks and Platforms: Langflow provides low-code visual workflows for multimodal and RAG-based multi-agent systems.Agents can process text and images, orchestrate interactions, and execute tasks without extensive coding.
  • Agent Frameworks and Platforms: SuperAGI focuses on rapid instantiation, lifecycle management, and scalable deployment of autonomous AI agents.The platform emphasizes practical task execution, coordination, and management despite lacking a direct academic publication.
  • Agent Frameworks and Platforms: AutoGen develops complex applications through multi-agent conversations involving human interaction, LLM decisions, and external service invocation.Its workflows cover mathematics, coding, question answering, and operational research.
  • Autonomous AI Agent Applications: OpenHands targets production-level autonomous code generation, while ResearchGPT automates literature review, information extraction, summarization, and synthesis.These applications illustrate autonomous planning, coding, iterative refinement, and research workflows.

C. Domain-specific AI Agents

Domain-specific AI agents specialize autonomous capabilities for particular operational contexts. The section highlights cybersecurity, autonomous driving, tutoring, finance, medical support, and edge-network applications, including LAENet.

  • C. Domain-specific AI Agents: Domain-specific AI agents are autonomous systems designed for tasks unique to particular application areas.They use specialized domain knowledge and targeted capabilities to improve performance and practicality in those contexts.
  • C. Domain-specific AI Agents: XAgent provides autonomous real-time network threat detection, security analysis, and dynamic cybersecurity responses.The framework is presented as an OpenBMB open-source project.
  • C. Domain-specific AI Agents: CARLA supports autonomous-driving research through sensor fusion, realistic urban simulation, and evaluations involving dynamic obstacles.Its scenarios include other vehicles and pedestrians, supporting training for self-driving delivery tasks.
  • C. Domain-specific AI Agents: VoxPoser uses language models and composable 3D value maps to interpret natural-language instructions for robotic manipulation.The supplied passage describes it as an adaptive personalized tutoring agent for these tasks.
  • C. Domain-specific AI Agents: Representative edge applications include LAENet, where Agentic AI combines LLM-based multimodal understanding and reasoning with reinforcement learning for adaptive aerial-network decisions.The UAV-assisted IoT scenario optimizes data collection and energy delivery under throughput, reliability, freshness, and energy constraints.

3) Workflow of Agentic AI framework for LAENet:

The LAENet Agentic AI workflow integrates compact LLMs with reinforcement learning across perception, action selection, reward feedback, and policy refinement. LLM-generated adaptive rewards improve energy outcomes and offer potential flexibility for more complex optimization tasks.

  • Workflow of Agentic AI framework for LAENet: Agentic AI combines LLM contextual comprehension and structured reasoning with RL’s sequential decision-making for LAENet optimization.The workflow is organized into four stages for multimodal, dynamic decision-making.
  • State Perception and Abstraction: State perception abstracts UAV location, residual energy, and channel conditions using lightweight LoRA-adapted or quantized LLMs.These compact models encode heterogeneous multimodal sensory signals and user instructions for edge deployment.
  • Action Selection and Policy Execution: Action selection uses chain-of-thought reasoning and causal anticipation to decompose objectives into subgoals and generate adaptive trajectories.Compact actor-critic or distilled policy modules support policy execution.
  • Reward Evaluation and Feedback Processing: Adaptive reward shaping interprets explicit and implicit feedback to balance data freshness with resource consumption under changing conditions.The approach accommodates environmental variability and user preferences better than manually designed functions.
  • Numerical Results: 6.4% lower final energy consumption is achieved by Agentic TD3 versus its manually designed counterpart.The LLM-generated reward incorporates UAV positional information and energy factors, reducing flight distances and communication overhead.
  • Lessons Learned: LLM-generated rewards show generalization potential for multi-objective and cross-domain resource allocation, while manually crafted rewards remain less flexible.The reported advantage is linked to richer, context-aware reward signals and action selection.

1) Background and Motivation:

Agentic contextual retrieval addresses the rigidity of centralized or rule-based retrieval and intent translation in dynamic edge networks. By combining compact LLMs, distributed memory, RAG, and multi-hop reasoning, it supports adaptive intent fulfillment with improved accuracy and communication efficiency.

  • Background and Motivation: Traditional retrieval architectures use centralized indexing or fixed rule-based matching, limiting scalability and responsiveness as network states change.The challenge is especially relevant to vehicular networks, aerial relays, and multi-agent systems.
  • Background and Motivation: Agentic retrieval interprets natural-language queries, reasons over task histories, and proactively retrieves or generates content aligned with mission objectives.This dynamic in-situ process differs from static retrieval frameworks.
  • Framework: The ACR framework equips network agents with compact LLMs, multi-hop communication, contextual prompts, and RAG access to distributed knowledge bases.Agents use these resources for semantic understanding, contextual reasoning, and adaptive intent interpretation.
  • ACR Workflow: ACR converts high-level intents into structured queries, retrieves and filters contextually relevant knowledge, performs multi-hop reasoning, and incorporates execution feedback.The workflow updates memory and future retrievals to maintain alignment with evolving task realities.
  • Numerical Results: 14.8% higher task success and 23.4% lower communication are reported for ACR under complex multi-agent intents and compared with traditional retrieval or uniform broadcast.The gains are attributed to semantic interpretation, distributed-memory reasoning, iterative prompt refinement, and context-aware knowledge prioritization.

5) Lessons Learned:

Agentic AI extends vehicular edge computing from static scheduling toward intent-responsive, semantically informed control. Vehicles combine visual semantic abstraction, policy retrieval or online decision-making, QoE-aware evaluation, and continual policy refinement.

  • Lessons Learned: Static scheduling and centralized decision logic struggle with mobility, variable wireless links, and heterogeneous users in vehicular MEC.These limitations become more pronounced in ultra-dense networks and semantically rich applications.
  • Lessons Learned: Vehicular agents parse natural-language objectives, perceive semantic environment signals, and coordinate offloading or scheduling decisions dynamically.The framework uses semantic inference and onboard AI processing within edge nodes.
  • System Description: The system models vehicles as embodied agents using cameras, LLAVA-based semantic extraction, V2I and V2V links, and base-station spectrum coordination.Vehicles operate across W subbands while extracting information such as object descriptions and parking availability.
  • Workflow: Vehicles convert visual observations into structured semantic representations, retrieve successful policies or use GAE-PPO, and generate adaptive transmission parameters.The parameters include transmission power, communication channel, and semantic symbol length.
  • Workflow: Semantic fidelity is evaluated with cosine similarity between BERT embeddings, while a Weber-Fechner-inspired QoE reward combines semantic accuracy and transmission cost.GAE-PPO updates the policy and stores performance traces for future retrieval.
  • Lessons Learned: The resulting design supports multimodal reasoning, human-perceived utility alignment, and real-time behavioral refinement for scalable semantic edge intelligence.These capabilities transform scheduling into an intent-responsive process.

4) Numerical Results:

Agentic AI frameworks combine semantic reasoning, preference modeling, and reinforcement learning to support adaptive, human-centric edge service provisioning. In vehicular networks, the Agentic AI-enabled method improves convergence and accumulated return over baseline policies.

  • Vehicular networks: 61% higher accumulated return than pure PPO demonstrates faster convergence and improved sample efficiency for Agentic AI-enabled vehicular decision-making.The method compares against pure PPO, DDPG, and a random policy.
  • Human-centric service provisioning: Human-centric service provisioning targets subjective preferences that conventional edge optimization objectives such as latency and throughput often overlook.The framework addresses personalized, context-aware service requirements rather than only generic system metrics.
  • Human-centric service provisioning: The framework uses LLMs to interpret natural-language preferences, convert them into preference vectors, and guide dynamic SFC optimization.Its architecture combines lightweight edge models with preference-guided decision making and feedback adaptation.
  • Human-centric service provisioning: A three-stage workflow interprets requests, composes preference-guided SFCs through DRL, and adapts using multimodal feedback and contextual memory.The workflow connects preference understanding, resource allocation, execution monitoring, and continual refinement.
  • Human-centric service provisioning: The integrated design shifts edge optimization from resource-centric objectives toward adaptive, human-centric provisioning while maintaining efficiency and scalability.The stated outcome is continual evolution toward individual user needs.

4) Numerical Results:

In personalized content generation, the proposed Agentic AI framework outperforms conventional optimization baselines by aligning service objectives with individual preferences. The survey concludes by identifying research directions for efficient, private, robust, transferable, and compression-aware agentic edge systems.

  • Numerical results: 27.3% improvement in human-centric QoE over traditional DRL methods with uniform user preferences demonstrates the value of preference-aware optimization.The evaluation covers technical reports, creative content, and data-analysis outputs with diverse quality expectations and contexts.
  • Numerical results: The framework achieves consistently superior performance across all reported evaluation metrics in personalized content generation services.The scenario includes requests with diverse quality expectations and contextual requirements.
  • Future directions: Future work emphasizes scalable decentralized collaboration, privacy-preserving federated agent systems, and robust safety mechanisms for autonomous reasoning.Proposed mechanisms include decentralized consensus, secure aggregation, hallucination detection, reasoning validation, and fail-safe operation.
  • Future directions: Cross-domain adaptation and knowledge migration are identified as directions for generalizing agentic systems across heterogeneous edge contexts under resource constraints.Suggested approaches include transfer techniques, self-supervised domain alignment, and continual learning.
  • Future directions: Compression-aware agentification research should co-design model compression with explicit reasoning to preserve responsiveness, cognitive expressiveness, and energy efficiency.Named techniques include low-rank adaptation, structured pruning, quantization, and knowledge distillation.
  • Survey scope: The survey reviews foundational concepts, enabling technologies, representative applications, research challenges, open-source platforms, and future directions for edge general intelligence.Its stated scope covers model compression, energy-aware computing, robust connectivity, and knowledge representation and reasoning.
Loading 2508.18725v1…