The CO2ST of Agentic AI in Research
CO2-стоимость агентного ИИ в научных исследованиях
2026-09-30
SCID: 54.1/7p9d84q3
Discuss with AI
academic literature reviewagentic AIcarbon footprintlarge language modelsmulti-agent systems
Figures from the paper
Abstract (AI)
Generative AI has been rapidly adopted into multiple research practices. In this paper, we focus on the newest trend, so-called agentic AI. We calculate the carbon footprint of using agentic AI in a simulation of an academic literature review. We show that the energy consumption increases proportionately with the level of agency, resulting in a carbon footprint about three times larger when moving from low agency to a multi-agent process. The most co2stly step of this system is processing large quantities of text by the LLM, and resource consumption grows proportionally to the model size. Finally, we show the energy consumption of the CPU and GPU during processing and illustrate that the use of LLMs versus other software largely drives the power consumption difference. We demonstrate that researchers may make environmentally sustainable choices in research practices by making deliberate choices about agentic AI use.
Key Findings
1
A simulation of an academic literature review quantified the carbon footprint of agentic AI use in research.
2
Deliberate choices about the level and configuration of agentic AI can help researchers reduce environmental impacts.
3
Energy consumption increased proportionately with the level of agency, producing an approximately threefold larger carbon footprint for a multi-agent process than for low agency.
4
LLM use, rather than other software, primarily drove differences in CPU and GPU power consumption during processing.
5
Processing large quantities of text by the LLM was the most carbon-intensive step, with resource consumption increasing proportionately with model size.
Research Object
agentic AI used in a simulated academic literature review
Research Subject
the energy consumption and carbon footprint of agentic AI, including their dependence on agency level, LLM size, text-processing workload, and LLM use
Publication Details
Publication Date
2026-09-30
Journal
Publisher
ISSN
Open access PDF
Access Type
Author Information
Download PDF
Subscribe to digest