Agentic AI-RAN: Enabling Intent-Driven, Explainable and Self-Evolving Open RAN Intelligence

This paper proposes an Agentic AI-RAN framework that integrates planning, tool use, memory, and self-management primitives into O-RAN controllers to enable safe, explainable, and self-evolving intent-driven operations, demonstrating an 8.83% reduction in resource usage across network slices compared to conventional ML/RL baselines.

The Gist

Imagine a massive, bustling city called Open RAN. This city is the backbone of our mobile internet, handling everything from your video calls to self-driving car data.

In the past, this city was run by a rigid set of traffic lights and manual switches. Recently, we upgraded it to be "smart," using AI to manage traffic. But this new AI had a problem: it was like a brilliant but short-sighted robot. It could react instantly to a traffic jam, but it couldn't plan ahead, it didn't remember past mistakes, and if it got confused, it couldn't explain why it turned a red light green. It was a "black box" that sometimes made dangerous decisions.

This paper proposes a new kind of manager for this city: Agentic AI-RAN. Think of this not as a robot, but as a highly skilled, thoughtful City Planner who has a team of specialized assistants.

Here is how this new system works, broken down into simple concepts:

1. The Three-Layer City Management

The city is managed by three different levels of planners, each with a different job and speed:

• The Chief Strategist (Non-RT RIC): This is the big-picture planner. They work slowly (every minute or so) and use a Large Language Model (LLM)—like a super-smart consultant. They look at the whole city, read the "intent" (e.g., "We need to make sure ambulances get through quickly without using too much fuel"), and create a long-term game plan.
• The Traffic Coordinators (Near-RT RIC): These are the mid-level managers. They take the Chief's plan and break it down into specific instructions. They are fast (sub-second) and make sure the plan fits within the budget and safety rules.
• The Street Workers (CU/DU): These are the actual workers on the ground. They execute tiny, split-second adjustments, like changing a traffic light for 10 milliseconds.

2. The "Toolbelt" of Skills

Instead of just guessing what to do, this new AI has a toolbelt of "Skills."

• Imagine the AI is a handyman. Instead of just "fixing traffic," it has specific tools: a "Power Cap" wrench, a "Handover" hammer, or a "Resource Allocation" screwdriver.
• The Magic: The AI doesn't just grab a tool and swing wildly. It plans which tools to use, in what order, and checks if it's safe before using them. If a tool doesn't work, it has a "rollback" button to undo the change immediately.

3. The "Memory" and "Conscience"

This is the biggest upgrade. The old AI had no memory; it lived only in the present. The new Agentic AI has:

• A Diary (Memory): It remembers what happened last week. If a certain traffic pattern caused a crash, it remembers that and avoids repeating the mistake.
• A Conscience (Guardrails): Before making a move, the AI asks itself: "Is this safe? Do I have enough battery? Will this break the rules?" If the answer is "No," it stops. It won't risk the whole city just to save a few seconds.
• A Notebook (Evidence): Every time it makes a decision, it writes down why it did it. This is crucial for auditors (the city inspectors) to see that the AI is acting fairly and safely.

4. The "Plan-Act-Observe-Reflect" Loop

Think of this as the AI's daily routine:

1. Plan: "I need to move 500 cars from the north to the south."
2. Act: "I will send a small group of cars first to test the road."
3. Observe: "The road is clear, but the traffic is slowing down slightly."
4. Reflect: "Okay, the plan is working, but I need to slow down the next group to be safe."

5. The Results: A Smoother Ride

The authors tested this new system in a simulation of a busy city. Here is what they found:

• Better Efficiency: The city used 8.8% less energy (fewer cars idling, less fuel wasted).
• Fewer Accidents: The "SLA violations" (dropped calls or slow internet) went down significantly.
• Transparency: Unlike the old "black box" AI, you can now ask the new AI, "Why did you slow down that traffic?" and it can show you its notes and explain its reasoning.

The Big Picture

This paper argues that we shouldn't just build faster AI; we should build smarter, more responsible AI. By giving the AI a plan, a memory, a set of tools, and a conscience, we can manage complex networks like Open RAN safely, efficiently, and in a way that humans can trust and understand.

It's the difference between a reckless race car driver who just speeds up and hopes for the best, and a professional fleet manager who plans routes, checks the weather, monitors fuel, and keeps a log of every trip.

Technical Summary

1. Problem Statement

The evolution toward Open RAN (O-RAN) has introduced flexibility and vendor neutrality but has created significant operational challenges:

• Complexity in Multi-Tenant Environments: Managing multi-objective, multi-tenant networks safely and auditably is difficult with current modular xApps/rApps.
• Limitations of Existing AI/ML: Traditional Machine Learning (ML), Reinforcement Learning (RL), and Multi-Agent RL (MARL) controllers in O-RAN often function as "black boxes." They typically rely on fixed task definitions, single objectives, and extensive offline training. They lack:
  • Explicit goal reasoning and planning.
  • Reusable skills and memory for cross-episode learning.
  • Self-governance mechanisms (safety guardrails) to handle uncertainty or conflicting objectives.
  • Explainability and auditability for regulatory compliance.
• The Gap: There is a need for a control architecture that moves beyond reactive optimization to intent-driven, self-evolving, and explainable control loops that span Non-RT, Near-RT, and Real-Time (RT) layers.

2. Methodology: The Agentic AI-RAN Framework

The authors propose an Agentic AI-RAN perspective where O-RAN control entities are modeled as goal-driven agents. These agents operate across the O-RAN architecture (Non-RT RIC/SMO, Near-RT RIC, and CU/DU) using a structured set of primitives.

Core Architecture & Primitives

The framework introduces four key agentic primitives:

1. Plan-Act-Observe-Reflect (PAOR):

   • Plan: Agents synthesize short control plans based on operator intents (grounded into machine-checkable goals) and current network context.
   • Act: Execution is performed via "skills" (O-RAN controllable actions) bound to standard interfaces (A1, E2, E2SM).
   • Observe: Agents monitor outcomes via active telemetry (KPM/KQI).
   • Reflect: Agents evaluate actions against risk, uncertainty, and budgets before proceeding.
   • Timing: Long-horizon planning occurs at the Non-RT layer (using Digital Twins and potentially LLMs), while short-horizon sequencing and gating happen at Near-RT and RT layers.

2. Skills as Tool-Use:

   • Control actions are encapsulated as "skills"—structured, verifiable wrappers with preconditions, expected effects, costs, and rollback options.
   • This allows for composable, incremental interventions rather than monolithic state changes, ensuring vendor-agnostic execution.

3. Memory & Evidence:

   • Memory: Multi-timescale storage including short-term state (Near-RT), episodic traces of decisions/outcomes, and long-term knowledge (Non-RT) for reusing verified policies.
   • Evidence: Decision-level records link intents, context, selected skills, and outcomes. This supports human/regulatory audits without exposing raw customer data.

4. Self-Management (Gating):

   • Agents enforce safety invariants via gates. Before executing a step, the agent checks:
     • Predicted SLA violation risk ($r_t$).
     • Uncertainty scores ($u_t$).
     • Resource budget usage ($b_t$).
     • Explanation consistency ($c_t$).
   • If a gate fails, the agent shrinks the step size, waits, or rolls back to a vetted baseline, ensuring graceful degradation.

Role of Large Language Models (LLMs)

• LLMs are strategically placed only at the Non-RT RIC/SMO due to strict real-time constraints in the RAN.
• They are used for intent grounding (translating natural language operator goals into structured machine goals), long-horizon reasoning, and summarizing Near-RT behaviors to optimize slice admission and resource allocation.

3. Key Contributions

• Conceptual Shift: Proposes treating O-RAN control entities as goal-driven agents rather than static ML models, enabling explicit planning, tool use, and self-management.
• Structured Framework: Formalizes the PAOR loop and defines agentic primitives (skills, memory, gating) that align with existing O-RAN interfaces (A1/E2) and time scales.
• Task Landscape Definition: Organizes O-RAN control tasks into three clusters:
  1. Network Slice Life-cycle: Design, admission, scaling, and troubleshooting.
  2. Radio Resource Management (RRM): Closed-loop control for PRB, power, and handover.
  3. Security, Privacy, & Compliance: Embedded as explicit constraints and budgets within the decision process.
• Comparative Analysis: Contrasts Agentic AI with traditional ML/RL/MARL, highlighting advantages in safety, auditability, data efficiency (via retrieval), and handling non-stationary environments.

4. Simulation Results

The authors evaluated the framework using a multi-cell O-RAN simulation (6 O-RUs, 12 cells, 20 UEs) comparing a Fully Agentic Controller against ablated variants and a conventional Deep Q-Learning (DQL) baseline.

• Performance Metrics:
  • SLA Violations & Latency: The Fully Agentic controller achieved the lowest SLA violations and p99 tail latency compared to all other variants.
  • Trade-offs: While the agentic approach incurred slightly higher operational overhead (E2 bytes, control actions) due to active telemetry and incremental adjustments, it significantly outperformed the baseline in Quality of Service (QoS).
  • Ablation Insights:
    • Removing planning or memory led to increased SLA violations.
    • Removing the gate caused a pronounced rise in instability and SLA costs.
    • Disabling active KPM reduced overhead but increased risk.
• LLM Impact: Integrating an LLM at the Non-RT layer (using ChatGPT 5.2) for slice coordination resulted in:
  • 5.7% to 8.3% increase in slice admission accuracy.
  • 7.1% to 10.3% reduction in resource usage.
  • 14.3% to 17.9% reduction in p99 latency.
  • Conclusion: LLMs improve resource efficiency and slicing precision when confined to non-real-time layers.

5. Significance and Future Directions

• Significance: This work bridges the gap between theoretical AI agents and practical O-RAN deployment. It demonstrates that explainable, safe, and intent-driven control is achievable without violating O-RAN timing constraints. It shifts the paradigm from "black-box optimization" to "transparent, auditable control."
• Future Work:
  • Integrating richer RL/MARL components within the agentic framework.
  • Testing in more realistic Digital Twin environments.
  • Developing standardized hooks (skill catalogues, evidence buses) for scalable deployment in operational networks.

In summary, the paper presents a robust architectural blueprint for Agentic AI-RAN, proving that combining planning, memory, and self-governance primitives can significantly enhance the safety, efficiency, and adaptability of next-generation Open RAN networks.