The strong performance of AI agents across a wide range of tasks has spurred massive investment in agentic infrastructures, yet the cost of processing tokens is rising rapidly. Web agents automate tasks described in natural language by analyzing the user interface of web applications and interacting with it. This paper introduces OdoBot, a novel web‑agent architecture that builds a behavioral model of the target application from successful task‑execution demonstrations, enabling it to complete tasks with far fewer tokens while maintaining or improving success rates. Experiments on 45 tasks in the Canvas Learning Management System show that OdoBot uses 44% and 80% fewer tokens than two state‑of‑the‑art competitors (Agent‑E and WebVoyager) and surpasses WebVoyager in task success rate.
Review