Token Efficient Task Execution via Application Behavior Modeling for Web Agents

Alexandru Ianta; Eleni Stroulia

cs.AISep 11, 2026

Token Efficient Task Execution via Application Behavior Modeling for Web Agents

Authors: Alexandru Ianta, Eleni Stroulia

Organizations: Department of Computing Science University of Alberta Edmonton, Canada

Abstract

The strong performance of AI Agents across an impressive variety of tasks is driving an unprecedented investment in agentic infrastructures, however the cost of processing tokens is fast increasing. Web agents automate the execution of web-application tasks described in natural language, by analyzing the web-application's user interface (UI) and interacting with it. This work introduces OdoBot, a novel web-agent architecture that completes tasks at a fraction of the cost when compared to conventional web agents. This is achieved by leveraging a behavioral model of the underlying application constructed by analyzing successful task-execution demonstrations. Our experiments with 45 tasks on the Canvas Learning Management System (LMS) demonstrate that OdoBot uses 44% and 80% fewer tokens than two state-of-the-art competitor agents (Agent-E and WebVoyager), while also surpassing WebVoyager in terms of task success rate.

Figures & tables

Agent	Task SR	Year	LLM
Agent-E	73.1% [ 9 ]	2024	gpt-4-turbo*
WebVoyager	59.1% [ 10 ]	2024	gpt-4-1106-preview
WILBUR	52.6% [ 8 ]	2024	gpt-4-0125-preview
WebVoyager [Text-only]	44.3% [ 10 ]	2024	gpt-4-1106-preview
NNetNav	35.2% [ 17 ]	2025	Llama8B

TABLE I: Reported task success rate (SR) on the WebVoyager benchmark for different agents. *The exact snapshot for Agent E is not reported [ 9 ] .

TABLE II: Trajectory events (left) and their mapping to event vertices (right). Comparing the ‘properties’ and ‘fixed parameter’ columns shows that for two click events to map to the same event vertex they must share base paths, xpaths and text values.

Fig. 1: A sample graph model, generated from trajectories a6aa9079 and 85409a14, that diverge only in the last user action.

Fig. 2: An illustrated example of a (prefix, tag, suffix)-tuple dynamicXpath. The green rectangle highlights the prefix container element. The blue rectangles highlight repeated HTML sub-structures identified by matching children with the dynamicXpath’s tag. The red rectangles highlight the elements matching the suffix. A click event annotated with this dynamicXpath can be resolved by clicking either element 1 or 2, and a click event whose xpath is either that of element 1 or 2 will be identified as appertaining to that click event. The color-coded element xpaths are shown in the top right.

TABLE III: A table mapping actions in the graph to the instructions executed by OdoBot to resolve the action at task execution time. Click* refers to click actions annotated with dynamicXpaths.

Agent	Success Rate $\%$	Total Tokens $\mathrm{m}\mathrm{i}\mathrm{l}\mathrm{l}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{s}$	Tokens/Task $\mathrm{t}\mathrm{h}\mathrm{o}\mathrm{u}\mathrm{s}\mathrm{a}\mathrm{n}\mathrm{d}\mathrm{s}$
OdoBot	$76.4\pm 2.0$	$0.9\pm 0.02$	$19.9\pm 0.4$
Agent-E	$85.3\pm 2.5$	$1.6\pm 0.10$	$35.2\pm 2.3$
WebVoyager [1920x1080]	$72.9\pm 1.0$	$4.6\pm 0.07$	$101.5\pm 1.6$
WebVoyager [1280x720]	$67.6\pm 3.4$	$3.2\pm 0.09$	$69.9\pm 2.1$
WebVoyager [854x480]	$34.7\pm 2.0$	$2.5\pm 0.05$	$56.3\pm 1.1$

TABLE IV: Mean task success rate, total token use, and token use per task for each agent for the 45 evaluation tasks over five runs.

Web Agents

Web agents use language and vision models to navigate websites, gather information, and complete tasks through interactive interfaces.

See all 179 papers

Read More, Think More: Revisiting Observation Reduction for Web Agents
Apr 2, 2026Masafumi Enomoto, Ryoma Obara, Haochen Zhang +1Web Agents
NEC Corporation
MolmoWeb: Open Visual Web Agent and Open Data for the Open Web
Apr 9, 2026Tanmay Gupta, Piper Wolters, Zixian Ma +13Web Agents LLM Agent Training
Allen Institute for AI · University of Washington · UNC-Chapel Hill
Back to Basics: A Skill-Centric Approach to Generalist Web Agents
Oct 26, 2026Minh Pham-Dinh, Md Farhan Ishmam, Hernan Yupanqui +2Web Agents LLM Agent Skill Learning
University of Utah · Islamic University of Technology

Apr 2, 2026J/K move · Enter open · S save

Read More, Think More: Revisiting Observation Reduction for Web Agents

Masafumi Enomoto, Ryoma Obara, Haochen Zhang, Masafumi Oyamada

NEC Corporation

Web Agents

Open paper AskPDF arXiv

Web agents based on large language models (LLMs) rely on observations of web pages -- commonly represented as HTML -- as the basis for identifying available actions and planning subsequent steps. Prior work has treated the verbosity of HTML as an obstacle to performance and adopted observation reduction as a standard practice. We revisit this trend and demonstrate that the optimal observation representation depends on model capability and thinking token budget: (1) compact observations (accessibility trees) are preferable for lower-capability models, while detailed observations (HTML) are advantageous for higher-capability models; moreover, increasing thinking tokens further amplifies the benefit of HTML. (2) Our error analysis suggests that higher-capability models exploit layout information in HTML for better action grounding, while lower-capability models suffer from increased hallucination under longer inputs. We also find that incorporating observation history improves performance across most models and settings, and a diff-based representation offers a token-efficient alternative. Based on these findings, we suggest practical guidelines: adaptively select observation representations based on model capability and thinking token budget, and incorporate observation history using diff-based representations.

Abstract

Web Agents

Explore similar work

Read More, Think More: Revisiting Observation Reduction for Web Agents

MolmoWeb: Open Visual Web Agent and Open Data for the Open Web

Back to Basics: A Skill-Centric Approach to Generalist Web Agents