eess.SYApr 9, 2026
SaveNetworking-Aware Energy Efficiency in Agentic AI Inference: A Survey
Organizations: Shanghai University, Shanghai, China · Nanyang Technological University, Singapore · Edith Cowan University, Perth, Australia · Fudan University, Shanghai, China ·
Abstract
The rapid emergence of Large Language Models (LLMs) has catalyzed Agentic artificial intelligence (AI), autonomous systems integrating perception, reasoning, and action into closed-loop pipelines for continuous adaptation. While unlocking transformative applications in mobile edge computing, autonomous systems, and next-generation wireless networks, this paradigm creates fundamental energy challenges through iterative inference and persistent data exchange. Unlike traditional AI where bottlenecks are computational Floating Point Operations (FLOPs), Agentic AI faces compounding computational and communication energy costs. In this survey, we propose an energy accounting framework identifying computational and communication costs across the Perception-Reasoning-Action cycle. We establish a unified taxonomy spanning model simplification, computation control, input and attention optimization, and hardware-aware inference. We explore cross-layer co-design strategies jointly optimizing model parameters, wireless transmissions, and edge resources. Finally, we identify open challenges of federated green learning, carbon-aware agency, 6th generation mobile communication (6G)-native Agentic AI, and self-sustaining systems, providing a roadmap for scalable autonomous intelligence.
Figures & tables
| Ref. | Overview | Agentic AI | Edge Inference | Wireless Networks | Energy- Efficiency |
|---|---|---|---|---|---|
| ( Vaidhyanathan and Taibi, 2026 ) | A comparative study evaluating the functional capabilities and reasoning accuracy of general-purpose Agentic AI frameworks (e.g., AutoGPT), but lacking insights into the energy constraints or communication overhead required for real-world wireless edge deployment. | ✓ | ✗ | ✗ | ✗ |
| ( Usman et al., 2025 ) | A detailed article proposing an edge-native intelligence framework for 6G networks that integrates AI/ML for autonomous network management and self-healing, addressing edge deployment constraints but without a specific focus on the generative agentic workflow. | Partially | ✓ | ✓ | ✗ |
| ( B et al., 2025 ) | A survey reviewing advanced quantization techniques (GPTQ, GGUF, AWQ) for Large Language Models (LLMs) within autonomous driving and ADAS, focusing on memory reduction and inference acceleration on resource-limited edge hardware while analyzing accuracy trade-offs. | ✗ | ✓ | ✗ | ✓ |
| ( Bhardwaj et al., 2024 ) | A survey delving into the integration of Large Language Models (LLMs) with edge computing, highlighting advantages in privacy and latency while addressing challenges in computational demand and energy efficiency for resource-limited device deployment. | ✗ | ✓ | ✓ | Partially |
| ( Zhang et al., 2026 ) | A comprehensive survey on Agentic AI frameworks for edge intelligence, introducing enabling technologies, representative case studies, and future directions toward scalable and trustworthy deployments in next-generation wireless edge networks. | ✓ | ✓ | ✓ | ✗ |
| ( Dantas et al., 2025 ) | A comprehensive review of state-of-the-art LLM compression techniques (e.g., pruning, quantization) aimed at lowering inference energy consumption, yet without addressing the communication overhead or distributed coordination inherent in wireless agentic networks. | ✗ | ✓ | ✗ | ✓ |
Table 1. Summary of Related Works
Figure 1. Survey organization: Section 2 introduces Agentic AI concepts and energy accounting; Section 3 reviews optimization methods; Section 4 examines cross-layer co-design for wireless-AI integration; Section 5 identifies open challenges and future directions; Section 6 provides concluding remarks.
| Dimension | DRL-Powered Agents | LLM-Powered Agents | Standard Agentic AI | Energy-Efficient Agentic AI |
|---|---|---|---|---|
| Core Driver | Policy Networks | Static LLMs | Perception-Reasoning-Action Loop | Resource-Aware Co-Design |
| Paradigm | Trial-and-Error Learning | Probabilistic Generation | Autonomous Planning | Adaptive Computation and Communication |
| Reasoning | Implicit / Weak | Static CoT | Dynamic Long-Horizon | Elastic Depth with Early Exit |
| Memory | Short-term Markov State | Fixed Context Window | Long-term Vector Database | Compressed KV Cache |
| Energy Profile | Very Low (Inference) | High (Linear Growth) | Extremely High (Multiplicative) | Controllable via Algorithm-Hardware Co-Design |
| Limitation | Poor Generalization | Lack of Autonomy | High Cost & Latency | Accuracy Trade-off |
Table 2. Evolutionary Chain of Intelligent Agents: From Specialized Optimization to Sustainable Autonomy
| Module | Primary Operation | Compute Energy | Memory Energy | Communication Energy | Key Bottleneck |
|---|---|---|---|---|---|
| Perception | Vision encoding, RAG retrieval | High | Medium | Low-Medium | Vision encoder FLOPs |
| Reasoning | LLM inference, CoT generation | Medium | Very High | Low-Medium | KV cache capacity |
| Action | Tool/API execution, coordination | Medium | Low | High | Multi-agent communication |
| Memory | KV cache management, retrieval | Low | Very High | Medium | Cache fragmentation |
Table 3. Energy Consumption Characteristics by Agentic AI Module
Figure 2. Generic framework illustrating how Edge Agentic AI autonomously integrates multimodal observations, scheduling-driven constraints, and compressed foundation models with hardware-aware API selection for the continuous Perception–Thinking–Reasoning–Action process in resource-constrained wireless networks.
Figure 3. An overview of energy efficient optimization methods.
| Technique | Ref. | Energy Saving Mechanism | Benefits | Limitation |
|---|---|---|---|---|
| Quantization | ( Husom et al., 2025 ) | Reduces precision to 3-bit or 4-bit formats. | Reduces energy consumption by up to 79% compared to FP16. | Accuracy drops 20-30% on complex reasoning tasks (e.g., GSM8K). |
| ( Hu et al., 2025 ) | Quantization-adaptive resource scheduling (QLLMS). | Reduces GPU rental costs by 22.36%; improves task completion by 59%. | Optimization complexity increases with resource heterogeneity. | |
| ( Zhang et al., 2025c ) | Joint batching and quantization allocation. | Enables real-time, privacy-preserving LLM services on edge. | Resource allocation becomes NP-hard with many users. | |
| Pruning | ( Xia et al., 2024 ) | End-to-end structured pruning (Sheared LLaMA). | Compresses 7B models to 1.3B/2.7B to fit edge budgets. | Requires retraining/fine-tuning to recover performance. |
| ( Frantar and Alistarh, 2023 ) | One-shot pruning (SparseGPT) for massive models. | Achieves 50-60% sparsity without extensive retraining. | Smaller models (e.g., 1.3B) incur noticeable accuracy drops. | |
| ( Tian et al., 2024 ) | Generative pruning-ratio model guided by SWaP constraints. | Reduces energy and latency by over 30% with stable perplexity. | Dependency-aware pruning adds pipeline complexity. |
Table 4. Summary of Model Simplification Techniques (Section 3.1 )
| Technique | Ref. | Energy Saving Mechanism | Benefits | Limitation |
|---|---|---|---|---|
| Token Length Control | ( Wei et al., 2025 ) | Learns concise token representations via Generative Information Bottleneck (GenIB). | Reduces computational complexity and communication energy. | Minimal accuracy loss vs. efficiency gain. |
| ( Zhang et al., 2025b ) | Task-oriented multimodal token transmission with sliding-window pooling. | Balances transmission latency against model accuracy. | Leads to 2.3% performance degradation in low-SNR regimes. | |
| ( Yang et al., 2024a ) | Enforces optimal maximum token limits based on queuing theory. | Reduces queuing delay and user drop rates. | Risks truncating necessary reasoning steps (context loss). | |
| Early Exit / Adaptive Depth | ( Jin and Wu, 2025 ) | Partitions LLMs: high-confidence local exit, low-confidence offload to cloud. | Avoids unnecessary full-model inference and transmission. | Dependent on reliable confidence estimation. |
| ( Zheng et al., 2025b ) | Hierarchical early exit across end-edge-cloud. | Terminates inference once confidence is reached. | High scheduling complexity across heterogeneous devices. | |
| ( Venkatesha et al., 2025 ) | Speculative drafting on edge, verification on cloud with early exits. | 21% speedup in vision-language control tasks. | Speedup relies on draft acceptance rate and network stability. |
Table 5. Summary of Computation Control Techniques (Section 3.2 )
| Technique | Ref. | Energy Saving Mechanism | Benefits | Limitation |
|---|---|---|---|---|
| Token Pruning | ( Jiang et al., 2023 ) | Coarse-to-fine compression pruning redundant tokens (LLMLingua). | Reduces prompt size up to 20 ; lowers transmission/prefill energy. | Potential loss of fine-grained context details. |
| ( Wang et al., 2021 ) | Progressively discards “lazy” tokens based on cumulative attention. | Orders-of-magnitude energy efficiency improvement on mobile. | Cause loss of subtle logical nuances in complex reasoning tasks. | |
| ( Zhong et al., 2025 ) | Merges redundant visual tokens and prunes via PageRank (AIM). | Reduces FLOPs (6.8 ) and prefill time (8 ) with 99.7% accuracy. | May over-simplify visual features in texture-heavy scenes. | |
| Sparse Attention | ( Dao et al., 2022 ) | I/O-aware tiling and recomputation to minimize HBM access (FlashAttention). | Accelerates inference; reduces memory traffic via SRAM usage. | Dependent on GPU memory hierarchy and SRAM size. |
| ( Hadish et al., 2025 ) | Integrates FlashAttention and ProbAttention for smart grids. | Low memory complexity suitable for constrained resources. | Struggles with long-term seasonal shifts. | |
| ( Tanwar et al., 2025 ) | Sparse autoencoders with Grouped Query Attention (M7BCO). | Reduces energy consumption in WSNs (20.2 mJ vs 22.5 mJ). | Latency may exceed the real-time demands of fast-fading channels. |
Table 6. Summary of Input and Attention Optimization Techniques (Section 3.3 )
| Technique | Ref. | Energy Saving Mechanism | Benefits | Limitation |
|---|---|---|---|---|
| Precision Scheduling | ( Li et al., 2023 ) | Allocates precision based on layer sensitivity. | Ensures accuracy under memory budgets of edge devices. | Increases scheduling and workload balancing complexity. |
| ( Bajpai and Gupta, 2025 ) | Integrates mixed precision with structured pruning (EcoLLM). | Avoids reconstruction errors; efficient compression for edge. | Increase implementation and hardware mapping complexity. | |
| ( Cho et al., 2025 ) | Sub-4-bit inference with Dual-Mode Matrix Multiplication Unit. | 1.82 higher energy efficiency; 11.8 compression. | Performance relies on specific log-scale structured sparsity. | |
| DVFS | ( Zhang et al., 2025a ) | Unified CPU-GPU-Memory governor identifying optimal freq triplets (FUSE). | Reduces latency by 7%-36.8% under same energy budget. | Requires specialized hardware to achieve the claimed energy efficiency. |
| ( Patel et al., 2024 ) | Overlaps GPU power capping with inference phases (POLCA). | Deploys 30% more servers under same power budget. | Limits portability across different mobile SoC models. | |
| ( Kurma et al., 2025 ) | IAI-LLM driven CPU frequency scaling and RIS beamforming. | Minimizes cumulative overhead by 44.96% via task-aware DVFS. | Frequent frequency switching increases hardware signaling overhead. |
Table 7. Summary of Hardware-Aware Inference Techniques (Section 3.4 )
Figure 4. An overview of integrated wireless-edge intelligence for sustainable Agentic AI.
More in this topic
Energy-Efficient ML
Energy-efficient machine learning reduces the energy and environmental costs of training and inference while preserving useful model performance.
See all 249 papers