This thesis presents the design, implementation, and evaluation of AI agentic workflows using the LangGraph framework, with a strong focus on energy and computational perfor- mance analysis. The research addresses the increasing demand for sustainable AI systems by integrating CodeCarbon for real-time energy consumption tracking, JMeter for concur- rency and response time testing, and Arize Phoenix for workflow traceability and quality evaluation. The developed system enables modular, parallel execution of LLM-driven tasks (with Llama 3.2), allowing dynamic adaptation to user load. Experimental results, conducted on a high-performance virtualized environment, demonstrate that increasing concurrency improves throughput but also escalates energy consumption, revealing trade- offs between scalability and eco-sustainability. The findings underline the importance of detailed monitoring and traceability in developing environmentally conscious agentic sys- tems, paving the way for further optimization and benchmarking in sustainable AI.
Questa tesi presenta la progettazione, l’implementazione e la valutazione di workflow agentici di intelligenza artificiale tramite il framework LangGraph, con particolare at- tenzione all’analisi delle prestazioni energetiche e computazionali. La ricerca risponde alla crescente esigenza di sistemi AI sostenibili integrando CodeCarbon per il monitor- aggio in tempo reale del consumo energetico, JMeter per i test di concorrenza e tempo di risposta, e Arize Phoenix per la tracciabilità e valutazione della qualità dei workflow. Il sistema sviluppato consente l’esecuzione modulare e parallela di task basati su LLM (Llama 3.2), adattandosi dinamicamente al carico utente. I risultati sperimentali, condotti su un ambiente virtualizzato ad alte prestazioni, mostrano che l’aumento della concor- renza migliora il throughput ma comporta anche un incremento del consumo energetico, evidenziando i trade-off tra scalabilità e sostenibilità ambientale. Questi risultati sot- tolineano l’importanza di un monitoraggio e una tracciabilità dettagliati nello sviluppo di sistemi agentici rispettosi dell’ambiente, aprendo la strada a future ottimizzazioni e benchmarking nell’AI sostenibile.
Development and evaluation of AI agentic workflows in LangGraph: an energy and computational performance analysis
TORRADO GUZMAN, DAVID ALEJANDRO
2024/2025
Abstract
This thesis presents the design, implementation, and evaluation of AI agentic workflows using the LangGraph framework, with a strong focus on energy and computational perfor- mance analysis. The research addresses the increasing demand for sustainable AI systems by integrating CodeCarbon for real-time energy consumption tracking, JMeter for concur- rency and response time testing, and Arize Phoenix for workflow traceability and quality evaluation. The developed system enables modular, parallel execution of LLM-driven tasks (with Llama 3.2), allowing dynamic adaptation to user load. Experimental results, conducted on a high-performance virtualized environment, demonstrate that increasing concurrency improves throughput but also escalates energy consumption, revealing trade- offs between scalability and eco-sustainability. The findings underline the importance of detailed monitoring and traceability in developing environmentally conscious agentic sys- tems, paving the way for further optimization and benchmarking in sustainable AI.| File | Dimensione | Formato | |
|---|---|---|---|
|
Torrado_7_2025.pdf
accessibile in internet per tutti
Descrizione: Text of the thesis.
Dimensione
6.37 MB
Formato
Adobe PDF
|
6.37 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/240243