ns3-agent: Fostering Integrated Perception-Communication-Computing Research for Agentic AI Services via Cross-Platform Co-Simulations

Abstract

The rapid proliferation of Generative AI (GenAI) has catalyzed the emergence of autonomous agentic AI services, spanning large language model (LLM) or vision-language model (VLM) based digital agents to vision-language-action (VLA) based embodied intelligence. Consequently, tokens and multimodal data streams have emerged as fundamentally new traffic modalities carried by modern wireless networks, actively reshaping next-generation evolutionary standards. Unlike conventional multimedia streaming, the end-to-end quality of experience of agentic AI services is governed by an intricate, bidirectional closed loop encompassing physical perception or request generation, wireless transmission, and edge inference computing. However, existing simulation platforms often treat network communication and generative intelligence as decoupled domains, failing to capture their dynamic, cross-layer interdependencies. To bridge this chasm, this paper introduces ns3-agent, a cross-platform co-simulation framework tailored for generic agentic AI workflows. By incorporating GenAI client environments, edge inference servers, and the ns-3 engine into a unified multi-tier architecture, ns3-agent orchestrates seamless cross-domain orchestration. To guarantee stringent temporal causality and high-performance data exchange, we implement a conservative time synchronization mechanism and shared-memory inter-process communication. Finally, we present a case study that couples a high-fidelity robotic physics engine with a live-deployed, operational Octo VLA foundation model over congested 802.11ax Wi-Fi environments. Simulation results reveal that wireless link quality severely degrades task-oriented performance metrics through complex temporal delays and semantic-transport layer coupling loops. Motivated by these observations, ns3-agent can serve as a reliable and foundational simulation sandbox to foster research in integrated perception-communication-computing for agentic AI services.