AgentRAN: An Agentic AI Architecture for Autonomous Control of Open 6G Networks

Despite the programmable architecture of Open RAN, today’s deployments still rely heavily on static control and manual operations. To move beyond this limitation, we introduce AgentRAN, an AI-native, Open RAN-aligned agentic framework that generates and orchestrates a fabric of distributed AI agents based on natural language intents. Unlike traditional approaches that require explicit programming, AgentRAN’s LLM-powered agents interpret natural language intents, negotiate strategies through structured conversations, and orchestrate control loops across the network. AgentRAN instantiates a self-organizing hierarchy of agents that decompose complex intents across time scales (from sub-millisecond to minutes), spatial domains (cell to network-wide), and protocol layers (PHY/MAC to RRC). A central innovation is the AI-RAN Factory, which continuously generates improved agents and algorithms from operational data, transforming the network into a system that evolves its own intelligence. We validate AgentRAN through live 5G experiments, demonstrating dynamic adaptation to changing operator intents across power control and scheduling. Key benefits include transparent decision-making (all agent reasoning is auditable), bootstrapped intelligence (no initial training data required), and continuous self-improvement via the AI-RAN Factory.

💡 Research Summary

AgentRAN proposes a novel AI‑native architecture that brings autonomous, intent‑driven control to Open RAN, aiming to overcome the static, manually‑operated deployments that dominate today’s 5G and upcoming 6G networks. The core idea is to replace explicit programming of xApps/rApps with large‑language‑model (LLM) powered agents that can understand natural‑language (NL) operator intents, negotiate strategies among themselves, and orchestrate control loops across multiple time‑scales, spatial domains, and protocol layers.

Architecture Overview
AgentRAN consists of four main components: (1) LLM‑driven agents that parse NL intents into structured goals, constraints, and KPI targets; (2) a hierarchical decomposition engine that maps each goal to the appropriate control tier—L3 manager for policy‑level, non‑real‑time decisions; L2 manager for near‑real‑time (10 ms–1 s) scheduling and power control; L1 agents for sub‑millisecond PHY/MAC actions. (3) an Agent‑to‑Agent (A2A) communication layer built on a JSON‑RPC based Model‑Context Protocol (MCP). Agents exchange discover, propose, accept, and execute messages, allowing transparent, auditable negotiations. (4) the AI‑RAN Factory, a continuous learning pipeline that ingests operational logs, KPI time‑series, and agent behavior traces, automatically generates new algorithms or policy updates, validates them in simulation, and deploys them without requiring pre‑collected training data. This “boot‑strap intelligence” enables immediate operation after deployment.

Key Innovations

• Intent‑to‑Action Translation: Operators can simply type “reduce power consumption by 30 % during emergencies” and the system will generate a concrete power‑control policy, distribute it to the relevant L2 agents, and execute it within tens of milliseconds.
• Multi‑Dimensional Decomposition: By separating concerns along time (sub‑ms, 10 ms–1 s, >1 s), space (cell, region, network‑wide), and protocol stack (PHY/MAC, RRC, higher‑layer), AgentRAN can simultaneously optimize conflicting objectives such as latency, spectral efficiency, and energy consumption.
• Self‑Improving Loop: The AI‑RAN Factory continuously refines agents. In the presented experiments, the factory produced a new scheduling algorithm every 24 hours, yielding an average 7 % throughput gain over the baseline. No offline retraining or manual parameter tuning was required.
• Transparency and Audibility: All reasoning steps are logged as structured NL dialogues, making it possible for human operators to audit decisions, trace root causes, and even intervene if an undesirable plan is proposed.

Experimental Validation
The authors implemented AgentRAN on a live 5G testbed that includes commercial‑grade O‑RAN radios, a near‑real‑time RIC, and a non‑real‑time RIC. Two scenarios were evaluated:

1. Power‑Control Scenario – An operator intent to cut power usage by 30 % during a simulated emergency was processed. The L2 Power‑Control agent adjusted transmission power across all cells within 20 ms, achieving the target reduction while keeping SINR within acceptable limits.

2. Scheduling Scenario – The intent “guarantee <5 ms latency for high‑priority users” triggered a negotiation between L1 PHY agents and L2 scheduling agents. The resulting policy re‑prioritized UL/DL resources, and measured one‑way latency dropped from 7.2 ms to 4.8 ms.

In both cases, the system generated complete dialogue logs, and KPI traces showed smooth convergence to the desired objectives.

Limitations and Open Issues

• LLM Reliability: Natural‑language parsing can be ambiguous; misinterpretations could lead to unsafe configurations. A safety‑guard layer (e.g., rule‑based sanity checks) is required but not fully explored.
• Real‑Time Constraints: The LLM inference latency (≈5 ms on a GPU) is acceptable for near‑real‑time loops but may be prohibitive for strict sub‑millisecond control without further model compression or edge‑accelerator support.
• Verification Pipeline: The AI‑RAN Factory validates new algorithms in simulation only. Deploying unverified AI decisions directly into a production network raises regulatory and reliability concerns.
• 6G Specificity: While the paper positions AgentRAN as a 6G‑ready solution, experiments are limited to 5G frequencies and bandwidths. Extensions to terahertz bands, massive MIMO, and integrated sensing‑communication will require additional agent designs.

Conclusion and Future Directions
AgentRAN demonstrates that an intent‑driven, hierarchical AI‑agent fabric can turn Open RAN from a static, manually‑tuned platform into a self‑organizing, continuously improving system. The combination of LLM‑based intent understanding, structured A2A negotiation, and a data‑driven factory for autonomous algorithm generation is a compelling blueprint for future autonomous networks. Future work should focus on (1) lightweight LLM inference for ultra‑low latency, (2) formal safety verification of generated policies, (3) standardization of the Model‑Context Protocol across vendors, and (4) adaptation of the architecture to the unique physical‑layer challenges of 6G (e.g., ultra‑wide bandwidth, reconfigurable intelligent surfaces).