From Intents to Actions: Agentic AI in Autonomous Networks

Telecommunication networks are increasingly expected to operate autonomously while supporting heterogeneous services with diverse and often conflicting intents – that is, performance objectives, constraints, and requirements specific to each service. However, transforming high-level intents – such as ultra-low latency, high throughput, or energy efficiency – into concrete control actions (i.e., low-level actuator commands) remains beyond the capability of existing heuristic approaches. This work introduces an Agentic AI system for intent-driven autonomous networks, structured around three specialized agents. A supervisory interpreter agent, powered by language models, performs both lexical parsing of intents into executable optimization templates and cognitive refinement based on feedback, constraint feasibility, and evolving network conditions. An optimizer agent converts these templates into tractable optimization problems, analyzes trade-offs, and derives preferences across objectives. Lastly, a preference-driven controller agent, based on multi-objective reinforcement learning, leverages these preferences to operate near the Pareto frontier of network performance that best satisfies the original intent. Collectively, these agents enable networks to autonomously interpret, reason over, adapt to, and act upon diverse intents and network conditions in a scalable manner.

💡 Research Summary

**
The paper presents a comprehensive Agentic AI framework designed to bridge the gap between high‑level service intents and low‑level control actions in modern radio access networks (RAN). Recognizing that existing heuristic and rule‑based methods cannot reliably translate natural‑language intents such as “ultra‑low latency”, “high throughput”, or “energy efficiency” into concrete actuator commands, the authors propose a three‑agent architecture: an Interpreter Agent, an Optimizer Agent, and a Preference‑Driven Controller Agent.

The Interpreter Agent leverages large language models (LLMs) to parse intents and convert them into structured Optimization Template Models (OTMs). To respect the compute and memory constraints of edge RAN equipment, the design splits the task between two small language models (SLMs): one dedicated to translation (the “Translator”) and another to in‑context reasoning (the “Advisor”). A sliding‑window monitor continuously gathers network observations and feedback on intent fulfillment, allowing the interpreter to iteratively refine the OTM, detect infeasible constraints, and re‑issue corrected templates.

The Optimizer Agent receives the OTM and formulates a tractable optimization problem. It introduces a preference space that captures trade‑offs among multiple objectives (e.g., latency vs. energy). Using Bayesian optimization—specifically a novel Preference‑Aligned eXploration Bayesian Optimization (PAX‑BO) scheme—the agent models objective correlations with Gaussian Process priors and selects acquisition functions that balance exploration of the preference simplex with exploitation of promising regions. Preferences are dynamically updated in response to real‑time network conditions, ensuring that the resulting optimization problem remains aligned with the operator’s evolving goals.

The Controller Agent implements multi‑objective reinforcement learning (MORL) to generate control policies that operate near the Pareto frontier of network performance. The core contribution is Distributed Envelope Q‑Learning (D‑EQL), a scalable variant of Envelope Q‑Learning. D‑EQL decouples a centralized learner (maintaining a deep Q‑network) from multiple actors that each maintain sharded prioritized replay buffers. Actors explore different points of the preference simplex in parallel while sharing a single preference‑conditioned policy network. The learning update combines vector‑valued TD targets with a cosine‑stability loss to improve convergence, and employs hindsight preference relabeling to recycle past experience for more efficient preference‑space coverage.

The authors validate the full stack on a high‑fidelity 5G‑Advanced (5G‑A) simulator, focusing on an intent‑aware radio resource management (RRM) use case that integrates link adaptation (LA). Experiments compare the proposed system against (i) a traditional heuristic RRM pipeline, (ii) a single‑goal reinforcement‑learning LA agent, and (iii) state‑of‑the‑art MORL baselines. Results show:

• Intent fulfillment rates exceeding 95 % across heterogeneous service mixes.
• A 12 %–18 % improvement in aggregate utility when handling multiple, potentially conflicting service goals simultaneously.
• Adaptation latency to intent changes reduced to sub‑second (≈0.8 s), far faster than the several‑seconds latency of conventional pipelines.

These gains demonstrate that the three‑agent architecture can translate natural‑language intents into actionable policies in real time, even under the sub‑millisecond decision constraints of RRM. The paper also situates its contributions within ongoing 3GPP, Open‑RAN, and TM‑Forum standardization efforts, arguing that Agentic AI is poised to become a foundational technology for 6G intent‑driven autonomous networks.

In summary, the work advances the field by (1) introducing a dual‑SLM interpreter that balances linguistic capability with edge resource limits, (2) formulating a preference‑aware Bayesian optimizer that dynamically aligns multi‑objective trade‑offs with network state, (3) delivering a distributed MORL controller capable of learning a universal, preference‑conditioned policy at scale, and (4) providing extensive simulation evidence that the integrated system outperforms existing heuristic and RL approaches. The authors’ vision of a fully autonomous, intent‑centric RAN is thus substantiated both theoretically and empirically.