Linghao Zhang张凌豪
中文
Back to publications
1/3 The agentic AI control plane on the MAPE-K framework (Fig. 2).
2/3 Hierarchical agent-RL collaboration in the Plan phase (Fig. 3).
3/3 First page

IEEE Network2026Early AccessFirst author3 citations

Agentic AI for SAGIN Resource Management: Semantic Awareness, Orchestration, and Optimization

Linghao Zhang, Haitao Zhao, Bo Xu, Hongbo Zhu, Xianbin Wang

IEEE Network, pp. 1–8, 2026

IEEE Xplore

Abstract

Space-air-ground integrated networks (SAGIN) promise ubiquitous sixth-generation (6G) connectivity but face significant resource management challenges due to heterogeneous infrastructure, dynamic topologies, and stringent quality-of-service (QoS) requirements. Conventional model-driven approaches struggle with scalability and adaptability in such complex environments. This paper presents an agentic artificial intelligence (AI) framework for autonomous SAGIN resource management by embedding large language model (LLM)-based agents into a Monitor-Analyze-Plan-Execute-Knowledge (MAPE-K) control plane. The framework incorporates three specialized agents, namely semantic resource perceivers, intent-driven orchestrators, and adaptive learners, that collaborate through natural language reasoning to bridge the gap between operator intents and network execution. At the core of the framework is a hierarchical agent-reinforcement learning (RL) collaboration mechanism, wherein LLM-based orchestrators dynamically shape reward functions for RL agents based on semantic network conditions. Validation through unmanned aerial vehicle (UAV)-assisted AI-generated content (AIGC) service orchestration in energy-constrained scenarios demonstrates that LLM-driven reward shaping achieves 14% energy reduction and the lowest average service latency among all compared methods. This agentic paradigm offers a scalable pathway toward adaptive, AI-native 6G networks, capable of autonomously interpreting intents and adapting to dynamic environments.

Cite

Linghao Zhang, Haitao Zhao, Bo Xu, Hongbo Zhu, Xianbin Wang, “Agentic AI for SAGIN Resource Management: Semantic Awareness, Orchestration, and Optimization”, IEEE Network, pp. 1–8, 2026, doi: 10.1109/MNET.2026.3695611.

© IEEE. Figures shown for personal use.

Deep dive

LLMs plan at the slow timescale. RL schedules at the fast timescale.

An agentic AI control plane for space-air-ground resource management, built on a Monitor-Analyze-Plan-Execute-Knowledge loop.

What breaks

Satellites, drones and ground stations share radio, compute and battery across layers that differ widely in delay and power. Classical optimization assumes a tidy static model. Deep reinforcement learning adapts, but it is a black box that cannot tell an intent for emergency reliability from one for best-effort capacity. Recent work with language models mostly asks for a one-off answer on a static snapshot, so nothing watches the result or corrects it.

The idea

Divide the work by timescale. A language model reads the situation and rewrites the reward function of a fast reinforcement-learning agent. It does not make the fast decisions. LLM calls take a few seconds, well decoupled from an RL loop deciding at the millisecond level.

Plate 1 The control plane on top of a SAGIN: semantic resource perceivers, intent-driven orchestrators and adaptive learners around a shared knowledge store, reaching tools through the Model Context Protocol. Fig. 2 of the paper.

Walk the loop

The stages below use the paper’s own worked example, a UAV cluster running out of energy.

LLM agents · secondsMonitorAnalyzePlanExecuteRL agent · millisecondsKnowledge

Who

Telemetry, through the Model Context Protocol

Collects telemetry and performance indicators from satellites, UAVs and ground stations, and normalizes them.

In the paper’s worked example

UAV-1 is down to 20% energy. UAV-2 still has 80%.

Inside the Plan phase

Two mechanisms connect the orchestrator to the RL agent. Reward shaping turns a semantic condition into a modified reward: under energy constraints, penalize high-power transmissions and favor energy-efficient paths. Diffusion-based exploration uses historical trajectories in the knowledge base to guide what the agent tries.

Plate 2 Hierarchical agent-RL collaboration. The orchestrator interprets intents and semantic states, then configures RL agents through reward shaping and diffusion-based exploration. Fig. 3 of the paper.

Keeping it safe

A language model in a control loop needs a guard. Every LLM output is checked against a JSON schema and parameter bounds, and the RL agents fall back to a safe default reward whenever the check fails. The learner updates prompts and rewards only when KPIs drift past a threshold or on a schedule, and each update carries a recorded reasoning trace.

How it was tested

The paper validates the mechanism in a simulated UAV-assisted AIGC orchestration scenario under energy constraints: 3 LEO satellites, 5 UAVs, 2 ground base stations and 50 concurrent tasks, comparing six methods. The comparison is in the paper.

100%

↑↓ Move↵ Open