Project GAINS

(Generative AI for Network Sustainability)

Deliverable D1.1

State of the Art Report on Agent Architectures Based on Large Language Model (LLM) and Its Energy Impact

Giordano, A. – McGann, O. – Fruncillo, D. – Haider, B. | 2025-11

Abstract

This work synthesizes the state of the art on agent architectures based on Large Language Models (LLM) applied to Network Monitoring & Management (NMM) in converging infrastructures (access, edge, and cloud) and frames its energy impact. Starting from the most recent literature on the use of LLMs in networking, telecom-domain models, and emerging forms of autonomous and intent-driven management, the report delineates relevant use cases for operations, configuration, change management, and security, highlighting the main failure modes (intent ambiguity, configuration synthesis errors, drift, and version misalignment) and the necessary mechanisms to achieve industrial reliability. The analysis integrates patterns of Retrieval-Augmented Generation (RAG) and hybrid retrieval, with an emphasis on strong metadata filters and source provenance to ensure grounding and auditability. In terms of energy, the literature converges on an operational point: in multi-step workflows, consumption increases primarily with calls and tokens (prefill/decode), thus the study prioritizes measurable optimizations on calls, context, and retries. These findings are concise and depend on runtime and hardware with an impact verification through PoC with measures per request and per token. In terms of energy, the document consolidates evidence that, in typical continuous loads of NMM and multi-step agent systems, the dominant drivers of the footprint are the number of calls to the model and the length of the context, rather than just the model size; this leads to the priority of measurable optimizations based on context reduction, routing, and model cascading, speculative decoding, and observability per phase (prefill/decode) and per token (Joule/token). Finally, the deliverable proposes a replicable experimental setup with retrieval metrics (Recall@k, MRR, nDCG), pipeline metrics (LLM calls, tokens, retries, escalations), and attributable energy measures, and presents a comparative evaluation of the main open-source technologies and relevant licenses for industrialization (RAG frameworks, vector databases, workflow/agent builders, tool integration standards, and permissive/copy-left/fair-code licenses).

Table of Contents

1. Introduction
2. Methods
2.1 Setting up the analysis and selecting evidence
2.2 Criteria for Maturity and Suitability
2.3 Replicable experimental design: datasets, metrics and energy measurements 
3. Results
3.1 Use Case Taxonomy and Downstream Tasks for NMM
3.2 Architectural evolution: from "LLM that responds" to "LLM that orchestrates"
3.3 RAG and retrieval in technical domains: hybrid, strong filters and provenance
3.4 Energy impact: dominant drivers and measurable optimizations
3.5 Integral analysis of open-source projects and emerging technologies 
3.5.1 Standards and frameworks for integration and orchestration
3.5.2 RAG frameworks and agent builders
3.5.3 Vector database and hybrid retrieval with strong metadata
3.5.4 Agent-oriented platforms oriented towards data governance and control  
3.5.5 Emerging technologies: ecosystems, licenses and use of constraints
3.5.6 Licensing Implications: Transferability and Compliance
3.6 Strategic directions: "sovereign" agents and AI factories
4. Discussion
5. Conclusion
Bibliography

Keywords

Agent LLM, network automation, telemetry, energy measurement, Sustainability KPI, osservability

Back to Public Abstracts Request full paper

Reference: GAINS deliverables are linked to Work Packages and milestones. For additional materials, visit Resources.