Abstract:

Current Open Radio Access Network (O-RAN) near- real-time RIC applications (xApps) rely on reactive, single- objective machine learning, struggling with conflicting net- work goals. We propose Intent-Networking, a cognitive control framework using lightweight Belief–Desire–Intention–eXtended (BDIx) agents as xApps that translate high-level operator intents into verifiable, explainable plans via E2SM-RC, integrating symbolic reasoning with pluggable Machine Learning (ML) for proactive arbitration. We provide a standards-compliant reference architecture, formal model, and novel algorithms with a containerized prototype (open-sourcing in progress). On a high-fidelity OAI-based emulator, our multi-agent BDIx xApp reduces ultra-Reliable Low-Latency Communication (uRLLC) p99 latency by 23 % (8.1 → 6.2 ms) and average radio power by 17 % (125 → 104 W) versus state-of-the-art reactive ML xApps, while improving inter-slice fairness (Jain: 0.81 → 0.94) and achieving a DIF explainability score of 0.9 at 2.5 % CPU/cell.