Energy Measurement of LLM-Based Agentic Systems

Design and validate a methodology to measure and analyze energy use of autonomous LLM agent workflows.

MSc Thesis Wageningen University & Research Open

LLM agents can solve complex tasks by chaining tools, memory, and reasoning loops. This thesis asks a practical question: how much energy do those loops cost, and where exactly is it spent?

The Problem

LLM-based agentic systems are rapidly being adopted for software engineering, analysis, and decision support. Compared with single-shot prompts, agents often execute long trajectories: planning, invoking tools, revising outputs, and calling models repeatedly.

These systems can deliver strong task performance, but their energy and carbon cost remain poorly characterized. Existing benchmarks mainly focus on quality, latency, or token usage, and provide limited insight into end-to-end energy consumption at the workflow level.

What This Thesis Is About

This thesis focuses on creating a reproducible methodology for measuring energy use in LLM-based agentic systems. The objective is to move beyond proxy metrics and provide evidence-backed analysis of where energy is consumed across model inference, tool execution, orchestration overhead, and retries.

International Collaboration

This thesis is carried out in collaboration with international colleagues in Italy. Students will benefit from an international research setting.

Objectives

  1. Define a measurement protocol for agentic workflows, including task scenarios, instrumentation boundaries, and reporting units.
  2. Instrument the full execution pipeline to capture compute and energy signals across LLM calls, tools, and orchestration logic.
  3. Compare agent design choices (models, planning depth, tool-use strategies, retry policies) and quantify trade-offs between quality, latency, and energy.
  4. Deliver practical guidance for building more energy-efficient agentic AI systems in real development settings.

What You Will Do

Phase Work
Literature & setup Review Green AI, LLM systems, and energy-measurement methods; select an agent framework and representative tasks
Method design Define metrics, workloads, and controlled experiment settings for fair comparisons
Implementation Build instrumentation and logging pipeline for agent runs and associated energy signals
Evaluation Run experiments across configurations; analyze energy-quality-latency trade-offs
Reporting Produce thesis artifacts: methodology, dataset of runs, and recommendations

Who Should Apply

Required courses:

  • Programming in Python (INF-22306)
  • Machine Learning (FTE-35306)
  • Big Data (INF-33806)

Required skills & mindset:

  • Strong Python programming and data analysis skills
  • Interest in AI systems and sustainability
  • Comfort with experimentation, benchmarking, and reproducibility
  • Familiarity with LLM tooling is a plus

Key References

  • Verdecchia, R., Sallou, J., & Cruz, L. (2023). A systematic review of Green AI. WIREs Data Mining and Knowledge Discovery, 13(4), e1507. doi:10.1002/widm.1507
  • Henderson, P., Hu, J., Romoff, J., Brunskill, E., Jurafsky, D., & Pineau, J. (2020). Towards the systematic reporting of the energy and carbon footprints of machine learning. Journal of Machine Learning Research, 21(248), 1-43.
  • Lacoste, A., Luccioni, A., Schmidt, V., & Dandres, T. (2019). Quantifying the carbon emissions of machine learning. arXiv:1910.09700.
  • Strubell, E., Ganesh, A., & McCallum, A. (2019). Energy and policy considerations for deep learning in NLP. ACL 2019, 3645-3650.
  • Luccioni, A. S., Jernite, Y., & Strubell, E. (2024). Power Hungry Processing: Watts Driving the Cost of AI Deployment? FAccT ‘24. doi:10.1145/3630106.3658542
  • Kim, B., Choi, Y., Mei, H., et al. (2025). The ML.ENERGY Benchmark: Toward Automated Inference Energy Measurement and Optimization. arXiv:2505.06371. arXiv:2505.06371
  • Ni, Y., Wang, Y., Lu, M., et al. (2026). Budget-Constrained Agentic Large Language Models: Intention-Based Planning for Costly Tool Use. arXiv:2602.11541. arXiv:2602.11541
  • Wang, Z., Luo, X., Liu, H., et al. (2026). Spend Less, Reason Better: Budget-Aware Value Tree Search for LLM Agents. arXiv:2603.12634. arXiv:2603.12634
  • Tripathy, A., Harshit, C. P., & Vaidhyanathan, K. (2025). SWEnergy: An Empirical Study on Energy Efficiency in Agentic Issue Resolution Frameworks with SLMs. arXiv:2512.09543. arXiv:2512.09543

Supervisors

 June Sallou