Reasoning and Acting (ReAct) with Multimodal LLMs: A Framework for Intent Driven 6G Networks

6G wireless networks are poised to realise Self-X capabilities through the native integration of an AI-plane within the network architecture. This AI-plane will enable autonomous orchestration and management driven by user intent. Recent advances in Generative AI (GenAI) introduce powerful capabilities for such intent-based management, particularly through the representation of multimodal data in a shared embedding space. This enables the development of autonomous agents that not only translate intent but also manage network functions autonomously, both before and after deployment. In this paper, we propose a comprehensive framework for reasoning-and-acting (ReAct) agents powered by multimodal large language models (LLMs), facilitating seamless translation of user intent into network actions. The proposed architecture integrates multimodal retrieval-augmented generation (RAG), persona-driven reasoning, and function calling to orchestrate tasks across vision, text, and radio modalities. We examine how RAG influences reasoning in multimodal agents, how LLMs can be specialised for telecom applications, and which performance benchmarks should guide their design. A UAV-based coverage optimisation case study demonstrates the framework’s ability to leverage vision data to enhance average signal-to-noise ratio (SINR), yielding substantial reductions in outage probability and improvements in throughput under Rician fading channels. Analytical characterisation and practical implementation results are presented to benchmark performance.