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

Abstract

The Internet of Agents (IoA) not only provides a foundational architecture for collaborative reasoning among artificial intelligence (AI) agents but also brings new challenges to task offloading. Unlike traditional task offloading, agentic AI tasks involve sustained multistep reasoning, planning, and context maintenance, thereby imposing coupled pressure on memory occupancy, inference latency, and energy consumption. To address these issues, we propose a two-tier agentic AI task offloading framework within the IoA scenarios. In the first tier, resource-requester agents (RRAs), e.g., autonomous vehicles (AVs), offload inference-intensive tasks to nearby resource providers. Such resource providers can be mobile resource agents (MRAs), e.g., neighboring vehicles, and infrastructure resource agents (IRAs), e.g., roadside units (RSUs). In the second tier, overloaded IRAs further procure supplementary inference resources from aerial resource agents (ARAs), e.g., uncrewed aerial vehicles (UAVs), to handle overflow workloads caused by limited residual video RAM (VRAM) and key-value (KV)-cache accumulation. We formulate the first-tier pricing and offloading interaction as a Stackelberg game, where MRAs/IRAs act as leaders and RRAs act as followers. Moreover, we model the second-tier IRA-ARA resource trading as a double Dutch auction (DDA). To solve the proposed hierarchical model, we develop an iterative best-response algorithm for the first-tier Stackelberg game and a diffusion-based deep reinforcement learning (DRL) algorithm for the second-tier DDA process. Numerical results demonstrate that the proposed framework can improve agentic task-offloading efficiency, auction reward, and social welfare while reducing information-exchange cost.