Networking-Aware Energy Efficiency in Agentic AI Inference: A Survey

A new academic survey from a team of eight researchers, including Xiaojing Chen and Dusit Niyato, tackles the looming energy crisis posed by the next wave of AI: autonomous agents. Unlike single-query models like GPT-4, agentic AI systems (e.g., AI that can browse the web, execute code, or control devices) operate in continuous Perception-Reasoning-Action cycles. The paper argues this creates a fundamental new challenge where energy bottlenecks are no longer just computational (FLOPs) but are compounded by the persistent, iterative cost of data communication across networks, especially in mobile edge and IoT scenarios.

The survey's core contribution is a proposed energy accounting framework that breaks down costs across the entire agentic pipeline. It establishes a unified taxonomy for efficiency, covering techniques from model simplification and computation control to input/attention optimization and hardware-aware inference. Crucially, it advocates for cross-layer co-design—jointly optimizing the AI model's parameters, wireless transmission strategies, and distributed edge computing resources—as the path forward.

Finally, the researchers map a roadmap for sustainable autonomous intelligence, identifying critical open challenges. These include developing federated green learning techniques, creating carbon-aware agency where agents consider their energy footprint, designing 6G-native AI systems, and ultimately engineering self-sustaining systems. The work serves as a critical warning and technical guide for developers and infrastructure planners as agentic AI moves from research labs into widespread deployment.