Contract-based Agentic Intent Framework for Network Slicing in O-RAN

Intent-based networking aims to simplify network operation by translating operator intents into a collection of policies, configurations, and control actions. However, this translation process relies on heuristics and loose coupling. It often results in unpredictable behavior and ambiguous safety standards. This paper presents a Contract-based Agentic Intent Framework (CAIF) for the radio access network (RAN). The proposed framework employs a closed-loop agentic pipeline that systematically audits user objectives against formal RAN constraints prior to actuation. The proposed CAIF decouples probabilistic intent extraction from strictly governed policy execution to enable the enforcement of deterministic safety guarantees. We use network slicing as a representative use case to demonstrate the design flow and validate the effectiveness of the proposed approach on an O-RAN testbed. Experimental results show that the closed-loop agentic pipeline of the proposed CAIF can effectively eliminate harmful intent executions observed in direct-actuation baseline approaches.

💡 Research Summary

The paper addresses a critical gap in the deployment of Intent‑Based Networking (IBN) for 6G‑era radio access networks, specifically within the Open RAN (O‑RAN) architecture. While IBN promises to let operators express high‑level goals without manually configuring low‑level parameters, existing pipelines rely on heuristic mappings and loosely coupled components. This leads to unpredictable behavior, safety ambiguities, and the risk of violating Service Level Agreements (SLAs) when network slicing is dynamically re‑configured.

To solve these problems, the authors propose the Contract‑based Agentic Intent Framework (CAIF). CAIF introduces a closed‑loop, four‑stage pipeline that combines large language models (LLMs) with a formal “Intent Contract” defined by the TMF‑921 Intent Management API. The stages are:

1. Profiling Agent – parses multi‑turn natural‑language operator requests, extracts structured intent fields (slice identifier, KPI targets, etc.), and maps them to a JSON‑LD schema.
2. Evaluator Agent – checks completeness, internal consistency, and compliance with deterministic constraints (e.g., maximum throughput increase of 5 %, specific policy mechanisms). If any violation is detected, the evaluator triggers a refinement loop that asks the profiling agent to correct only the offending parts.
3. Contract Generator – converts the validated intent into a machine‑readable Intent Contract. The contract explicitly lists scope (icm:target), expectations (icm:hasExpectation), quantitative limits (Target Value), the policy mechanism to be used, and relationships to existing IntentSpecifications and policy baselines.
4. Intent Management – registers the contract, tracks its lifecycle, and performs conflict detection against already deployed policies before any control action is issued.

The key novelty lies in separating probabilistic intent extraction (handled by LLMs) from deterministic execution (enforced by the contract). The authors use two LLMs: a lightweight Qwen3‑4B‑Instruct model for profiling and a high‑capacity llama‑3.3‑nemotron‑super‑49b model for evaluation. Temperature is set to 0.6 and top‑p to 0.95 to limit randomness, while the contract validation layer guarantees that only semantically correct and policy‑compliant commands reach the RAN.

Experimental Setup
The framework is evaluated on an O‑RAN testbed that mirrors a realistic disaggregated RAN, using a commercial AI RSG RAN simulator for the radio layer. A curated dataset of 500 natural‑language intent instances is generated, covering 1‑shot to 5‑shot variations. Two baselines are compared: (a) a direct‑actuation pipeline where the LLM’s output is sent straight to the network, and (b) the proposed CAIF pipeline.

Results

• Translation Accuracy: The direct‑actuation baseline achieves 96.8 % accuracy (95 % confidence interval 94.9‑98.0 %). Sixteen errors are observed, twelve of which are false positives—syntactically valid A1 policies that do not reflect the operator’s intent, posing a severe safety risk. The CAIF pipeline reaches 99.8 % accuracy (95 % CI 98.9‑100.0 %), with only one semantic mismatch.
• Latency: CAIF introduces a modest overhead in the 1‑shot scenario (11.8 s vs. 8.5 s for the baseline) due to the two‑agent coordination. However, from the 2‑shot case onward, CAIF’s latency becomes lower than the baseline because the refinement loop quickly resolves ambiguities that would otherwise cause repeated re‑tries.
• Field‑Level Metrics: Across key intent attributes (slice ID, target throughput, policy type), CAIF achieves 100 % correctness with a 95 % CI of 99.2‑100 %, outperforming the baseline’s 95 %‑level performance.

Implementation Details
The contract is consumed by a hierarchy of O‑RAN components: an SLA Slice rApp translates percentage‑based requests into absolute throughput thresholds using live and historical PRB/CQI data from the O1 interface; an A1 Policy Handler rApp reads the contract and forwards the policy payload to the Near‑RT RIC via the A1 Mediator; three xApps (KPIMON‑GO, SLA Slice, and RC) form a closed‑loop control system that monitors KPI measurements, evaluates contract compliance, computes new resource‑ratio limits, and finally programs the radio through E2 messages. The contract’s flattened characteristic array (e.g., eligibleClusters) enables low‑latency access for the xApps without parsing complex graph structures.

Limitations and Future Work
The current contract schema focuses on throughput and resource‑ratio KPIs; extending it to cover latency, reliability, and multi‑objective QoS will be necessary for broader 6G use cases. The computational cost of large LLMs remains a bottleneck for real‑time operation; exploring model compression, distillation, or edge inference is an open direction. Finally, the framework currently lacks an automated rollback or safe‑mode mechanism for contract violations; integrating such fault‑tolerance features will be essential for production deployments.

Conclusion
CAIF demonstrates that coupling LLM‑driven intent interpretation with a rigorously validated contract can provide both the flexibility of natural‑language operators and the deterministic safety guarantees required for autonomous network slicing in O‑RAN. The experimental evidence shows a substantial reduction in semantic errors, improved SLA adherence, and acceptable latency overhead, marking a significant step toward trustworthy, intent‑driven automation in next‑generation mobile networks.