Towards Sustainable Web Agents: A Plea for Transparency and Dedicated Metrics for Energy Consumption
Lars Krupp, Daniel Geißler, Paul Lukowicz, Jakob Karolus · 2025
Read Towards Sustainable Web Agents: A Plea for Transparency and Dedicated Metrics for Energy Consumption by Lars Krupp, Daniel Geißler, Paul Lukowicz, Jakob Karolus (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-17903.txt.
Introduction
Lars Krupp and co-authors count the energy cost of web agents in this 2025 paper. Web agents are language models that browse and act online for users. They promise to absorb dull, repeated tasks, and research on them is booming.
Their upkeep, though, is rarely measured. The team gives a first look at the energy and carbon cost of running such agents. Design philosophies differ sharply in how much energy they burn. The authors plead for open reporting of those costs. They also call for standard metrics built for this purpose. Without them, users cannot weigh an agent's help against its footprint.
Catalog record
| CREATOR | Lars Krupp, Daniel Geißler, Paul Lukowicz, Jakob Karolus |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 24 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-17903
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Towards Sustainable Web Agents: A Plea for Transparency and Dedicated Metrics for Energy Consumption Lars Krupp, Daniel Geißler, Paul Lukowicz, Jakob Karolus Abstract Improvements in the area of large language models have shifted towards the construction of models capable of using external tools and interpreting their outputs. These so-called web agents have the ability to interact autonomously with the internet. This allows them to become powerful daily assistants handling time-consuming, repetitive tasks while supporting users in their daily activities. While web agent research is thriving, the sustainability aspect of this research direction remains largely unexplored. We provide an initial exploration of the energy and C O 2 CO_{2} cost associated with web agents. Our results show how different philosophies in web agent creation can severely impact the associated expended energy. We highlight lacking transparency regarding the disclosure of model parameters and processes used for some web agents as a limiting factor when estimating energy consumption. As such, our work advocates a change in thinking when evaluating web agents, warranting dedicated metrics for energy consumption and sustainability. Keywords Large Language Models \cdot Web Agents \cdot Sustainability 1 Introduction Web agents are LLM-powered systems capable of ”browsing the web”. Considered the next milestone in advancing large language models GoogleDeepMind 2025 ; David 2025 , the idea enables automated interaction with the internet. In essence, allowing web agents to explore much like a human would. While this concept has offered vast potential for intelligent tools, high computational costs Samsi et al. 2023 still remain a major issue. Context window sizes have exploded, exponentially increasing energy consumption and questioning the sustainability of this research direction. LASER Ma et al. 2023 is a popular web agent, able to achieve a success rate on the WebShop Yao et al. 2022 benchmark equivalent to the average success rate achieved by humans. Its capabilities include the ability to automatically find, select and buy a product on a website depending on a user query. The agent is capable to find, compare, select, and buy different articles. To do this, the agent presses buttons, uses search bars and navigates menus. While this autonomous interaction is impressive, so is its energy bill. An estimated 2930Wh per action (button press, using a search bar, …) on the Mind2Web Deng et al. 2024 benchmark. With this, using LASER Ma et al. 2023 emits approximately 9.6 kg CO_{2} per task on the Mind2Web benchmark. This is equivalent to a car driving 39 km (nearly the distance of a marathon). We argue that such approaches are not sustainable in the long run and advocate for a more holistic evaluation of web agent and in particular LLM performances, including energy spent to achieve benchmark results. In this work, we present an illustrative comparison between two popular web agents. MindAct Deng et al. 2024 , a web agent using open source LLMs and smart preprocessing to reduce its energy bill and the aforementioned LASER Ma et al. 2023 . By estimating and comparing the amount of energy consumed by these web agents with vastly different design philosophies, we highlight the impact of a web agent’s design on its energy consumption. Using a conservative estimation ( Section 3.1 ), LASER spends approximately 1500 times the energy than MindAct. Further, we point out the general difficulties with energy consumption estimation due to a lack of model parameter disclosure and propose which values should be reported in the future when releasing new web agents to make it possible to compare web agents based on their environmental impact. Our work encourages a change in thinking on how we evaluate web agent performance and disclosure of model parameters, reprimanding missing transparency. We show the stark differences in energy consumption between state-of-the-art web agents and propose metrics to report which allow a straightforward comparison of their energy efficiency. We advocate for a holistic evaluation of web agents incorporating dedicated metrics for energy consumption. 2 Related Work With the rapid improvements of LLMs in recent years Dubey et al. 2024 and their ever improving capabilities in tool-use Dubey et al. 2024 a new frontier of research has become possible. With the goal of building agents that can interact with the internet much like a human would, web agents recently are gaining traction Deng et al. 2024 ; Yao et al. 2022 . In this newly developing field, many different benchmarks are being proposed in rapid succession Yao et al. 2022 ; Deng et al. 2024 ; He et al. 2024 . While efforts to unify these benchmarks exist Chezelles et al. 2024 , at the moment, comparing the performance of different web agents is often not feasible. Analogously, the approaches in web agent construction show a high diversity. While some only use HTML for their input Ma et al. 2023 ; Deng et al. 2024 , others use the accessibility tree instead Chezelles et al. 2024 or supplement it with screenshots Zheng et al. 2024 or even use screenshots exclusively, like Pix2Act Lù et al. 2024 . This extends into the approaches used for preprocessing Deng et al. 2024 ; Gur et al. 2023 , the integration of memory modules Ma et al. 2023 and which kind of language models are used. While some approaches use open source models Deng et al. 2024 ; Gur et al. 2023 many use proprietary models Ma et al. 2023 ; Zheng et al. 2024 ; Yang et al. 2024 ; Zhang et al. 2024 making a precise evaluation of the environmental impact difficult. 2.1 Environmental Impact of LLMs There is a pressing discussion on sustainability and the environmental impact of developing and deploying LLMs Bender et al. 2021 . Since LLMs are trained on massive amounts of data to provide thorough knowledge, large-scale data centers are required to train the complex model architectures in sufficient time properly. Even though detailed information about LLMs are usually not publicly available, data from previous versions such as OpenAIs GPT-3 already state the usage of 175 billion model parameters being trained on 570 GB of data Brown 2020 . Statistics on the energy consumption are even sparser and commonly rely on rough estimations due to multiple unknown factors, such as the hardware architecture and efficiency, the training and optimization strategies, and most importantly the energy mix. In work by Lannelongue et al. 2021 , a Green Algorithm Calculator is proposed to estimate the Carbon Footprint of LLM training. Depending on the utilized hardware’s location, the energy mix between fossil and renewable energies severely impacts the environmental stress. For instance, 20g CO2e kWh (carbon dioxide equivalent per kilowatt hour) in Norway and Switzerland to over 800g CO2e kWh in Australia, South Africa, and the USA. Based on the BERT model Devlin 2018 , trained in the USA, they calculated a potential environmental impact of 0.754 metric tons of CO2 for a single training of 79h on 64 Tesla V100 GPUs with an average utilization of 62.7%. For the popular GPT-3 model, also based on the USA energy mix, it is estimated that around 550 metric tons of CO2 emissions were produced to complete the full training, exceeding the previous estimations from BERT tremendously due to the complexity and dataset size increase