Hybrid Stackelberg Game and Diffusion-based Auction for Two-tier Agentic AI Task Offloading in Internet of Agents

Yue Zhong, Yongju Tong, Jiawen Kang, Minghui Dai, Hong-Ning Dai, Zhou Su, Dusit NiyatoarXiv 2025 · DOI · arXiv

Abstract

The Internet of Agents (IoA) represents a paradigm shift where autonomous entities utilize large-scale AI models to interact with the physical world. However, the intensive computational requirements of these models often exceed the local capabilities of mobile agents. This paper presents a novel two-tier hierarchical offloading framework designed to mitigate these resource constraints. The first tier utilizes a multi-leader multi-follower Stackelberg game to manage task offloading between ground-based agents. The second tier employs a Double Dutch Auction (DDA) facilitated by a diffusion-based Deep Reinforcement Learning (DRL) algorithm to leverage aerial resources. Experimental results demonstrate that this hybrid approach significantly enhances social welfare and computational efficiency compared to traditional optimization and reinforcement learning baselines.

Introduction

As AI agents transition from isolated models to interconnected networks within the Internet of Agents (IoA), they face a critical "compute-mobility" paradox. Wireless Agents (WAs), such as autonomous vehicles, require split-second decision-making powered by massive sensor streams. Processing these tasks locally often leads to latency and overheating, yet offloading is complicated by resource scarcity and the need for fair incentive mechanisms in highly dynamic environments.

The multi-layered architecture of the Internet of Agents (IoA) system, illustrating the interactions between Wireless, Fixed, Mobile, and Aerial agents.

To address these challenges, the researchers propose a sophisticated hierarchical marketplace that allows resource-constrained agents to "borrow" brainpower from edge servers and aerial drones. This system categorizes agents into four types: Wireless Agents (WAs), Mobile Agents (MAs), Fixed Agents (FAs), and Aerial Agents (AAs), operating across distinct layers as illustrated in the system architecture above.

Methodology: A Two-Tier Offloading Framework

The proposed framework employs a hybrid approach to manage task offloading across different agent layers, ensuring economic fairness and computational efficiency.

The two-tier agentic task offloading mechanism combining Stackelberg Games and Double Dutch Auctions.

Tier 1: The Ground Game (Stackelberg Model)

The interaction between WAs and nearby resource providers (MAs and FAs) is modeled as a Multi-leader Multi-follower Stackelberg Game. In this tier:

  • Leaders (MAs/FAs): Set the price for their computing resources based on their own utility functions.
  • Followers (WAs): Determine their task offloading ratios based on the set prices and their local latency requirements.

An Iterative Best-Response algorithm is used to find the equilibrium where no agent can improve their utility by unilaterally changing their strategy, providing a stable pricing mechanism for ground-level offloading.

Tier 2: The Aerial Auction (Double Dutch Auction)

When ground-based Fixed Agents become overloaded, they can further offload tasks to Aerial Agents (AAs), such as UAVs, via a Double Dutch Auction (DDA). The primary innovation here is the use of a Diffusion-based DRL algorithm to manage the auction. Unlike standard DRL methods that may converge to local optima, the diffusion model learns the global distribution of optimal auction "clock steps." This allows the auctioneer to find the market-clearing price more rapidly and with lower communication overhead.

Experimental Evaluation

The framework was validated through numerical simulations grounded in real-world technical specifications, including heterogeneous processing capabilities and power coefficients.

Key Results

  • Superior Social Welfare: The diffusion-based DRL scheme achieved a social welfare score of approximately 178.17, significantly outperforming PPO (144.72), Greedy (124.44), and Random (112.10) baselines.
  • Economic Efficiency: The DDA mechanism was proven to be Incentive Compatible (IC) and Individually Rational (IR), ensuring agents are motivated to bid truthfully and are protected from financial loss.
  • Fast Convergence: The Iterative Best-Response algorithm for the Stackelberg game stabilized in just a few iterations, demonstrating its viability for real-time vehicular environments.

Conclusion and Future Work

This hybrid approach effectively addresses the compute-mobility paradox in the IoA by leveraging both game theory and generative AI. While the results are promising, future research will address limitations such as network disruptions due to high-speed mobility and the potential for malicious agents to manipulate auction prices. Expanding the model to include satellite-based edge computing could further enhance the reach of the Internet of Agents in remote environments.

Summary updated

Abstract

The Internet of Agents (IoA) is rapidly gaining prominence as a foundational architecture for interconnected intelligent systems, designed to facilitate seamless discovery, communication, and collaborative reasoning among a vast network of Artificial Intelligence (AI) agents. Powered by Large Language and Vision-Language Models, IoA enables the development of interactive, rational agents capable of complex cooperation, moving far beyond traditional isolated models. IoA involves physical entities, i.e., Wireless Agents (WAs) with limited onboard resources, which need to offload their compute-intensive agentic AI services to nearby servers. Such servers can be Mobile Agents (MAs), e.g., vehicle agents, or Fixed Agents (FAs), e.g., end-side units agents. Given their fixed geographical locations and stable connectivity, FAs can serve as reliable communication gateways and task aggregation points. This stability allows them to effectively coordinate with and offload to an Aerial Agent (AA) tier, which has an advantage not affordable for highly mobile MAs with dynamic connectivity limitations. As such, we propose a two-tier optimization approach. The first tier employs a multi-leader multi-follower Stackelberg game. In the game, MAs and FAs act as the leaders who set resource prices. WAs are the followers to determine task offloading ratios. However, when FAs become overloaded, they can further offload tasks to available aerial resources. Therefore, the second tier introduces a Double Dutch Auction model where overloaded FAs act as the buyers to request resources, and AAs serve as the sellers for resource provision. We then develop a diffusion-based Deep Reinforcement Learning algorithm to solve the model. Numerical results demonstrate the superiority of our proposed scheme in facilitating task offloading.
BibTeX
@misc{zhong2025hybrid,
  author = {Yue Zhong and Yongju Tong and Jiawen Kang and Minghui Dai and Hong-Ning Dai and Zhou Su and Dusit Niyato},
  title = {{Hybrid Stackelberg Game and Diffusion-based Auction for Two-tier Agentic AI Task Offloading in Internet of Agents}},
  year = {2025},
  eprint = {2511.22076},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.22076},
}
Figure