Multi-Agentic AI for Conflict-Aware rApp Policy Orchestration in Open RAN
A new AI system uses three specialized LLM agents to manage 5G networks, cutting reasoning costs by 95%.
A team of researchers from the University of Bristol and the University of Exeter has published a paper proposing a novel Multi-Agentic AI framework designed to solve a critical bottleneck in Open Radio Access Network (RAN) architecture. In Open RAN, which enables flexible, multi-vendor mobile networks, rApps generate strategic control policies while xApps handle real-time functions. Currently, rApp development is manual and brittle, struggling to scale as network complexity grows. The new framework automates this by deploying three specialized large language model (LLM) agents that work in concert: a Perception agent to analyze the network state, a Reasoning agent to synthesize control policies, and a Refinement agent to resolve conflicts and optimize decisions.
The system is augmented with retrieval-augmented generation (RAG) and memory-based analogical reasoning to improve accuracy and efficiency. In experiments across diverse deployment scenarios, the framework demonstrated a dramatic 70% improvement in deployment accuracy compared to baseline methods. More impressively, it achieved a 95% reduction in reasoning cost, a key metric for operational efficiency. The architecture also maintained zero-shot generalization, meaning it could handle new, unseen network intents without additional training. This establishes a scalable, conflict-aware path toward fully autonomous, "zero-touch" network orchestration, a major goal for telecom operators.
This research represents a significant step in applying multi-agent AI to real-world infrastructure problems. By moving policy generation from a manual, error-prone process to an automated, AI-driven one, it addresses a core challenge in the promised flexibility of Open RAN. The demonstrated cost and accuracy improvements could accelerate the adoption of disaggregated networks, paving the way for more intelligent, efficient, and self-optimizing 5G and future 6G systems.
- Uses three specialized LLM agents (Perception, Reasoning, Refinement) with RAG and memory to automate network policy generation.
- Achieved a 70% improvement in deployment accuracy and a 95% reduction in reasoning cost versus baseline methods.
- Enables zero-shot generalization to unseen network intents, supporting fully autonomous, "zero-touch" Open RAN orchestration.
Why It Matters
Automates the complex management of next-gen 5G/6G networks, reducing costs and errors for telecom operators.