- The paper introduces an agentic framework that decouples cognitive reasoning from deterministic actuation to refine ambiguous operator requests in 6G NaaS.
- The paper details a collaborative intent co-creation process using multi-agent cycles and a TM Forum-aligned knowledge graph to ensure precise order execution.
- The paper empirically evaluates the framework with open-source LLMs, showing that domain adaptation and catalog adherence outperform simple model scaling.
Agentic Intent Co-Creation for 6G Network-as-a-Service: Framework, Architecture, and Model Evaluation
Introduction and Motivation
The evolution toward 6G networks brings extreme service heterogeneity, multi-domain complexity, and demanding operational requirements for ultra-low latency, reliability, and sustainability. Manual or semi-automated approaches are insufficient for such environments, necessitating the shift to highly-autonomous, intent-based orchestration paradigms. However, traditional Intent-Based Networking (IBN) systems are limited, assuming well-formed intents and lacking mechanisms for handling ambiguous or incomplete operator requests. Post-provisioning assurance further exacerbates architectural mismatches. This paper introduces an agent-based, intent-driven end-to-end orchestration framework for 6G NaaS, operationalizing collaborative intent co-creation via a pool of domain expert agents and a TM Forum-aligned Body-of-Knowledge (BoK).
Architectural Contributions
Decoupled Cognition and Actuation
A fundamental principle in the proposed architecture is the decoupling of cognition (agentic reasoning powered by LLMs) from deterministic actuation (standardized execution controllers). Cognitive agents—Intent Co-Creation Agent and Domain Expert Agents—iteratively refine user intents, grounding them against catalog-backed knowledge bases before formalizing machine-readable orders. This segregation guarantees operational trust, ensures compliance with safety requirements, and mitigates risks associated with direct AI control in production-grade orchestration.
Collaborative Intent Co-Creation
Intent co-creation is modeled as a multi-agent iterative process, where the initial human or application request undergoes a cycle of enrichment (SLAs, QoS, cost, and business constraints), validation, decomposition into sub-goals, and domain-specific feasibility checks. The collaborative agents operate through an Agent-to-Agent messaging backbone, maintaining context and coherence via a dual-layer memory system: short-term collaborative working memory and long-term external case files. Once operational constraints and action plans are confirmed with the user, domain controllers execute deterministic workflows according to catalog-driven specifications.
TM Forum-aligned Body-of-Knowledge (BoK)
The BoK aggregates TM Forum information models, product offerings, service/resource specifications, site-specific constraints, and policy graphs in a hierarchical knowledge graph. This enables semantic mapping between high-level intents and commercial/technical layers, facilitating deterministic, traceable translation of user requirements into service and resource configurations. As a "single source of truth," it underpins both the co-creation process and runtime traceability for full lifecycle management.
Prototype Implementation
The prototype leverages ETSI OpenSlice as the operational OSS, interfacing with the BoK and implementing agentic orchestration via message-driven multi-agent microservices. Agents are deployed as Java Spring Boot services, subscribing asynchronously to JMS queues and invoking LLM inference backends (local via Ollama, or remote providers). The design maintains strict separation between cognitive reasoning and tool-mediated actuation through MCP-based access to OpenSlice APIs and workflows.
Agent Skill Integration
Agents are augmented with standardized skills, bridging LLM reasoning with operational tools and APIs. For example, Intent-to-Service Mapping skills parse natural language requests, extract constraints, query the BoK for relevant catalog products, and generate structured service plans. A policy enforces catalog adherence, explicit user confirmation, and avoidance of hallucinated products or orders, ensuring deterministic, standards-compliant orchestration.
Experimental Evaluation
Benchmark Scenario and Metrics
The agentic framework was evaluated against a ground-truth configuration provided by a senior 5G engineer, using a representative multi-domain NaaS intent encompassing latency, budget, coverage, and observability constraints. The evaluation focused on:
- Functional Composition: Recovery of the exact four-product bundle from the catalog.
- Constraint Satisfaction: Accurate budget computation and lifecycle specification.
- Technical Initialization: Generation of valid, catalog-aligned order payloads.
Metrics included product composition correctness, hallucinated products, cost accuracy, temporal reasoning, baseline achievement, and total dialogue time.
Open-Source LLM Comparison
Results highlighted significant disparities in model robustness and reliability. Gpt-oss-20b achieved the strictest instruction compliance, with zero hallucinations and proper handling of cost and dates; several larger or more popular models hallucinated product IDs, failed temporal reasoning, or violated explicit constraints.
Strong empirical finding: Model parameter count is not correlated with efficacy in domain-specific intent refinement; domain grounding and fine-tuning are more impactful for operational NaaS orchestration than raw scale.
Practical and Theoretical Implications
The framework demonstrates a practical solution for bridging the semantic gap between human-centric intent and machine-executable actions in 6G NaaS, using agentic reasoning explicitly isolated from deterministic execution. This approach achieves operational trust and safety required by production environments, supports ongoing quality assurance, and enables dynamic, multi-step negotiations of constraints and trade-offs.
Theoretically, the decoupled cognition-actuation model and collaborative intent co-creation establish a template for scalable and explainable AI-driven orchestration, aligning with industry standards (TM Forum, ETSI, 3GPP) and adapting to the rapid proliferation of new 6G service types and operational conditions.
Future Directions
Potential developments include:
- Optimizing agentic orchestration for edge environments with constrained compute footprints through memory-efficient inference.
- Comparative benchmarking against commercial LLMs (OpenAI, Anthropic, Mistral) for model selection in telco-grade scenarios.
- Extending the framework to encompass full service and resource orchestration beyond intent-to-order translation.
- Incorporating continuous learning and real-time conflict mitigation for greater network adaptability.
Conclusion
This paper formalizes an agentic intent-driven orchestration architecture for 6G NaaS, resolving critical shortcomings of static IBN by integrating collaborative co-creation, strict cognition-actuation separation, and TM Forum-aligned knowledge grounding. Empirical evaluation reveals domain adaptation as key to reliable order translation, emphasizing the importance of deterministic reference models and catalog adherence. The agentic framework provides a scalable foundation for autonomous, multi-domain orchestration and opens avenues for further research into explainable, trustworthy, AI-native networking systems.
