---
title: 'AgentRAN: Agentic Control for Open RAN'
url: https://www.emergentmind.com/topics/agentran
type: topic
---

# AgentRAN: Agentic Control for Open RAN

Searching arXiv for recent papers on AgentRAN and closely related agentic AI-RAN/Open RAN control frameworks.
AgentRAN denotes a family of agentic control architectures for Radio Access Networks in which high-level intents are interpreted by autonomous agents and translated into concrete RAN actions across the O-RAN and AI-RAN control stack. In the recent literature, the term appears both as the proper name of a specific Open 6G architecture based on distributed LLM-powered agents and as a broader label for agentic RAN control systems that combine intent interpretation, planning, tool use, memory, multi-agent coordination, and guarded execution for management, optimization, observability, and assurance [2508.17778, 2511.02532, 2602.13868, 2602.24115]. This suggests that “AgentRAN” has evolved from a single architectural proposal into a wider research program for intent-driven, explainable, and self-managing RAN intelligence.

## 1. Terminological scope and emergence

The 2025 paper "AgentRAN: An Agentic AI Architecture for Autonomous Control of Open 6G Networks" defines AgentRAN as an AI-native, Open RAN-aligned framework that generates and orchestrates a fabric of distributed AI agents based on natural-language intents. Its core claim is that Open RAN already provides the necessary interfaces and disaggregation, but lacks a semantic control plane that can bridge operator goals and low-level execution. The architecture therefore uses LLM-powered agents to interpret natural-language intents, negotiate strategies through structured conversations, and orchestrate control loops across time scales, spatial domains, and protocol layers [2508.17778].

A parallel but more conceptual line appears in "Agentic AI for Mobile Network RAN Management and Optimization," where AgentRAN is presented as a practical vision for applying Agentic AI to RAN operations. There, the emphasis is on moving beyond static scripts, fixed thresholds, and manually operated analytics toward a goal-driven autonomous system based on Large AI Models, KPI time-series analytics, memory, reflection, planning, and multi-agent coordination [2511.02532]. A related 2026 work, "Agentic Assistant for 6G: Turn-based Conversations for AI-RAN Hierarchical Co-Management," uses AgentRAN as the name of an agentic network manager and turn-based conversation assistant for AI-RAN hierarchical co-management, explicitly targeting co-management of RAN and native edge AI rather than one-shot RAG-based configuration advice [2602.13868].

The scope of the term broadens further in "Agentic AI-RAN: Enabling Intent-Driven, Explainable and Self-Evolving Open RAN Intelligence," which frames agentic RAN control as a long-lived controller layer above conventional ML/RL xApps. That article organizes the task landscape around three clusters: network slice life-cycle, radio resource management closed loops, and cross-cutting security, privacy, and compliance [2602.24115]. Across these works, AgentRAN is consistently associated with intent-driven control, explicit tool interfaces, and persistent operational context rather than isolated inference.

## 2. Architectural patterns and control hierarchy

A recurring structural principle is alignment with the O-RAN hierarchy. The original AgentRAN architecture distributes control across rApps in the non-real-time RIC, xApps in the near-real-time RIC, and dApps in the real-time domain. Agents are arranged in a recursive parent-child model: higher-level agents receive broader intents and broader context, while lower-level agents receive refined sub-intents and execute more specific actions. The architecture is explicitly described as spanning time scales from network-wide policy changes to fast scheduling loops, spatial domains from the whole network to specific cells, and protocol layers from service objectives down to PHY/MAC/RRC actions [2508.17778].

Subsequent systems refine this hierarchy into more explicit reasoning and execution roles. "Agentic Open RAN: A Deterministic and Auditable Framework for Intent-Driven Radio Control" introduces A1gent, in which a non-RT agentic rApp compiles operator goals into typed A1 policy instances and three task-oriented near-RT agentic xApps enforce them through a deterministic loop with plane-scoped actuation: E2 for mobility and load steering, and O1 for energy orchestration. The paper’s central architectural move is a strict split between non-RT reasoning and near-RT deterministic execution [2604.13384].

Other works expand the number of specialized agents. The 5G RAN case study in "Agentic AI for Mobile Network RAN Management and Optimization" defines a Master Orchestrator Agent, Analysis Agent, Historical Retrieval Agent, Documentation Agent, and Validation Agent operating over PM, FM, CM, IM, and historical sources [2511.02532]. "MX-AI: Agentic Observability and Control Platform for Open and AI-RAN" deploys a directed multi-agent graph in the SMO layer with Routing, Monitoring, Planner, Policy Synthesizer, Validator, and Executor functions over a live OAI/FlexRIC testbed [2508.09197]. "Agentic Assistant for 6G" organizes its assistant into an interface layer, an intelligence layer, and a knowledge layer centered on a Knowledge Router that dispatches queries to specialized knowledge sources covering user equipment, base stations, cells, RIC components, and AI services [2602.13868].

| System | Control split | Distinctive mechanism |
|---|---|---|
| AgentRAN | rApps, xApps, dApps | Recursive parent-child intent hierarchy |
| A1gent | non-RT rApp plus near-RT xApps | Typed A1 contracts and deterministic loops |
| MX-AI | SMO-layer agent graph | Observability and control on live OAI/FlexRIC |
| Agentic Assistant for 6G | interface, intelligence, knowledge layers | Turn-based hierarchical co-management |

Taken together, these designs show a convergent architecture: a slow-timescale intent and policy layer, a faster execution layer, and specialized agents for retrieval, validation, and state interpretation.

## 3. Intent translation, planning, and control semantics

AgentRAN research is centrally concerned with mapping abstract intents into bounded control parameters. In the original AgentRAN prototype, operators issue natural-language intents such as prioritizing MTC during emergencies or guaranteeing URLLC RTT below 20 ms. Top-level agents parse these goals, decompose them into sub-intents, negotiate feasibility with lower-layer agents, propagate constraints upward, and finally drive concrete scheduler and power-control actions through dApps. Agent-to-Agent communication is carried through A2A, while Model Context Protocol is used for discovering and calling APIs through natural language; both are stated to be built on JSON-RPC [2508.17778].

A more formal treatment appears in RIDAS, which is identified as an AgentRAN-style framework for AI-RAN because an intelligent agent actively controls RAN behavior by translating intent into low-level resource and model-configuration actions. RIDAS splits the system into representation-driven agents at the user end and an intention-driven agent at the base station. The intent translation step is written as
$$
\boldsymbol{\theta}_{N+1} = f_{\boldsymbol{\Phi}}\big(S, R, \{\boldsymbol{\theta}_i\}\big),
$$
with control parameter $\boldsymbol{\theta}=\{r,q\}$, where $r$ is the decomposition rank and $q$ is the number of quantization bits. The intention-driven agent uses a two-stage planning pipeline: bandwidth pre-allocation followed by bandwidth reallocation, so that new users are not immediately rejected when idle bandwidth is insufficient [2507.13140].

Intent translation is also formalized in cell-free O-RAN. In "Agentic AI for Intent-driven Optimization in Cell-free O-RAN," a supervisor agent deployed as an rApp converts the operator’s natural-language intent into an optimization objective and minimum rate constraints, then passes them to near-RT agents over A1. A user weighting agent retrieves prior experience from memory, an O-RU management agent is activated in energy-saving mode and uses multi-agent proximal policy optimization, and a monitoring agent coordinates the loop until minimum-rate requirements are satisfied [2602.22539]. Here the intent directly determines whether the utility is aggregate utility or energy saving, and whether the active O-RU set becomes a decision variable.

Negotiated intent is a second important semantic pattern. Agoran’s Service and Resource Broker receives stakeholder-side intents, generates Pareto-optimal offers with NSGA-II, and drives negotiation between stakeholder-side Negotiation Agents and an SRB-side Mediator Agent. The paper states that these agents reach a consensus intent in a single round, after which the agreed policy is pushed to Open and AI-RAN controllers [2508.09159]. This makes AgentRAN not only an operator-to-controller interface, but also a framework for reconciling multiple service owners and multiple objective functions.

## 4. Representative systems and empirical evidence

The literature now contains live testbeds, simulators, and emulators that report quantitative evidence for agentic RAN control. The results are heterogeneous because they target different layers—resource allocation, slice brokering, observability, placement, and assurance—but they consistently measure the agentic system against fixed baselines or ablations [2507.13140, 2508.09197, 2508.09159, 2602.22539, 2602.24115, 2605.07547].

| System | Evaluation setting | Representative reported result |
|---|---|---|
| RIDAS | AI-RAN image-classification resource allocation | Supports up to 123 users versus 85 for WirelessAgent; 44.71% more users |
| MX-AI | Live OAI/FlexRIC Open RAN testbed | 4.1/5.0 mean answer quality, 100% decision-action accuracy, 8.8 s latency |
| Agoran | Private 5G testbed with mobility traces | 37% throughput increase, 73% reduction in URLLC latency, 8.3% PRB saving |
| Cell-free O-RAN agentic AI | Simulation with shared LLM plus PEFT | 41.93% reduction in active O-RUs; 92% memory reduction |
| Agentic AI-RAN | Multi-cell O-RAN simulation | 8.83% average reduction in resource usage across three classic network slices |
| HAF | AI-RAN compute-sharing simulation | 90.0% overall SLO fulfillment; AI service request fulfillment from 51% to 85.3% |

RIDAS evaluates intent-driven control over CIFAR-10 using a ViT-B/16 backbone from CLIP as the representation model and DeepSeek-V3-0324 as the LLM. Under the same bandwidth budget, it supports up to 123 users, compared with 85 for WirelessAgent, and therefore supports 44.71% more users while reducing average per-user bandwidth and adapting allocation to latency tolerance more intelligently than the baselines [2507.13140].

MX-AI instruments a live 5G Open RAN testbed based on OpenAirInterface and FlexRIC and exposes observability and control functions through natural-language intents. Over 50 realistic operational queries, it reports mean answer quality of 4.1/5.0, 100% decision-action accuracy, and 8.8 seconds end-to-end latency with GPT-4.1 [2508.09197]. Agoran, deployed on a private 5G testbed with realistic vehicle-mobility traces, reports a 37% increase in throughput of eMBB slices, a 73% reduction in latency of URLLC slices, and an end-to-end 8.3% saving in PRB usage compared to a static baseline. The same paper further reports that a 1B-parameter Llama model fine-tuned for five minutes on 100 GPT-4 dialogues recovers approximately 80% of GPT-4.1’s decision quality, operates within 6 GiB of memory, and converges in only 1.3 seconds [2508.09159].

Evidence for specialized agentic control also appears outside generic Open RAN automation. The cell-free O-RAN framework reports a 41.93% reduction in the number of active O-RUs in energy-saving mode and a 92% reduction in memory usage when PEFT is used instead of deploying separate LLM agents [2602.22539]. The hierarchical agentic framework for AI-RAN compute sharing reports 90.0% overall SLO fulfillment, a 20.5% improvement over the strongest baseline, and an increase in AI service request fulfillment from 51% to 85.3% [2605.07547]. In the demonstrator article on Agentic AI-RAN, adding an LLM at the Non-RT layer improves admission accuracy and reduces resource usage and p99 latency for eMBB, URLLC, and mMTC slices, while the paper summarizes the overall result as an average 8.83% reduction in resource usage across three classic network slices [2602.24115].

## 5. Safety, assurance, and governance

A common misconception is that AgentRAN means direct free-form LLM actuation over the RAN. Recent work moves in the opposite direction: toward typed policies, deterministic execution, explicit safety gates, and auditable control paths. A1gent makes this design choice explicit by decoupling reasoning from real-time actuation. The non-RT rApp emits schema-checked A1 policy instances; the near-RT xApps enforce them in deterministic loops; and action conflicts are resolved by a fixed-priority merger
$$
\text{Energy} \triangleright \text{QoE} \triangleright \text{Load}.
$$
The same system adds hard constraints such as clamps, cooldown timers, budgets, minimum dwell times, and minimum active sectors, together with a training-free Adaptive Policy Tuner that only makes small bounded edits to soft policy fields [2604.13384].

The assurance-oriented architecture in "Agentic AI for Ultra-Modern Networks: Multi-Agent Framework for RAN Autonomy and Assurance" inserts independent verification between prediction and deployment. Its Orchestrator Agent, Data Collector Agent, Model Trainer Agent, Model Validator Agent, Predictor Agent, Policy Generator Agent, Simulator/Baseline Agent, Verifier Agent, Drift Detector Agent, Deployment Agent, Audit and Explainability Agent, and Security Agent turn a conventional predict-and-push pipeline into a predict, verify, explain, then deploy model. In the two-cell traffic-steering experiment, the no-agent mode improves the target cell locally but destabilizes the neighbor, whereas the agentic system blocks unsafe policies and preserves global network health [2510.16144].

The survey-style article on Agentic AI-RAN formalizes this safety logic as a self-management gate. At decision epoch $t$, the agent advances only if
$$
r_t \le \alpha,\;\; u_t \le \beta,\;\; b_t \le \mathcal{B},\;\; c_t \ge \tau,
$$
where $r_t$ is predicted SLA-violation risk, $u_t$ is uncertainty, $b_t$ is budget usage, and $c_t$ is explanation-consistency score. Otherwise, the controller must shrink the step, wait, or rollback [2602.24115]. This is complemented by governance mechanisms in Agoran, whose Legislative, Executive, and Judicial branches handle compliance grounding, live observability, and trust arbitration. Agoran’s Judicial branch evaluates each agent message with a rule-based Trust Score and applies real-time incentives to restore trust when malicious or manipulative behavior is detected [2508.09159].

At the same time, the empirical literature remains explicit about failure modes. "Agentic Assistant for 6G" reports an average response time of 13.35 seconds, planning fidelity and tool-use metrics that vary by task class, a planning-layer delegation accuracy of 48%, 41 documented cases of redundant planning steps, a tool success rate of 85%, and a hallucination rate of 43%. The paper emphasizes that correct tool use does not guarantee correct end-to-end reasoning [2602.13868]. MX-AI likewise states that human-in-the-loop review and stronger alignment mechanisms remain essential because LLMs can emit unsafe or incorrect control commands [2508.09197].

## 6. Relation to earlier RAN intelligence and open problems

AgentRAN did not emerge in a vacuum. Earlier multi-agent RAN work already distributed decision-making across the network edge, but typically without natural-language intent translation, persistent memory, explicit reflection, or typed governance. "A Multi-Agent Deep Reinforcement Learning Approach for RAN Resource Allocation in O-RAN" deploys local DDQN-based decision agents near the Near-RT RIC, each using local monitoring information from a shared gNodeB instance and coordinating through a shared reward and optional federated learning. The paper reports fair allocation, improved resource-utilization efficiency, and reduced over-provisioning risk on a real-time testbed [2307.02414]. "Intelligent Load Balancing and Resource Allocation in O-RAN: A Multi-Agent Multi-Armed Bandit Approach" makes each O-RU an agent and splits control between Non-RT RIC load balancing and Near-RT RIC resource allocation, using UCB-based multi-armed bandits rather than full RL [2303.14355].

The 2025 Agentic AI framework for RAN management makes the distinction explicit by contrasting Workflow, AI Agent, and Agentic AI. In that comparison, Agentic AI is characterized by high autonomy, unified planning, memory, reflection, tool interfaces, persistent memory, abstract goal handling, and dynamically constructed workflows [2511.02532]. This suggests that AgentRAN is best viewed not as “multi-agent RAN control” in the narrow sense, but as a shift from fixed pipelines and narrow policies to long-lived, goal-directed controllers that can decompose tasks, call tools, reason over multiple timescales, and preserve context.

The open problems identified across the literature are correspondingly architectural rather than purely algorithmic. The survey on Agentic AI-RAN points to richer RL/MARL integration, standardized skill catalogues, evidence buses, policy-as-code guardrails, interoperability across vendors, telemetry overhead, and timing constraints for standards-aligned deployment [2602.24115]. MX-AI asks how to reduce latency enough for tighter near-RT control loops, how to guarantee safety and alignment, how to scale from a small supported action set to richer network management, how to make retrieval robust at 6G telemetry scale, and how to standardize interoperable open ecosystems [2508.09197]. A1gent identifies hardware-in-the-loop E2/O1 integration and expansion of the A1 policy catalog as future work [2604.13384]. The general Agentic AI paper emphasizes computation and energy cost, security and privacy, communication and coordination, persistent memory, and interoperability as deployment challenges [2511.02532].

Across these trajectories, AgentRAN functions as an organizing concept for a new control plane in which intent parsing, planning, retrieval, bounded action, and assurance are treated as first-class RAN functions. The field’s central question is no longer whether agents can be inserted into the RAN stack, but how they can be made deterministic where necessary, adaptive where useful, and auditable throughout.

Source: https://www.emergentmind.com/topics/agentran