OdoBot Cuts Web-Agent Token Use Through Application Behavior Modeling

Summary

AI agents are increasingly used to execute web-application tasks from natural-language instructions, but the token cost of analyzing interfaces and taking actions is rising. The paper introduces OdoBot, a web-agent architecture that first builds a behavioral model of the underlying application by analyzing demonstrations of successful task execution. It then uses that model to complete tasks more efficiently than conventional agents that repeatedly reason over the application interface. In experiments covering 45 tasks in the Canvas Learning Management System, OdoBot used 44% fewer tokens than Agent-E and 80% fewer than WebVoyager. The reported reductions indicate that the architecture can lower processing costs substantially. OdoBot also achieved a higher task success rate than WebVoyager in the same evaluation, although the abstract does not provide the absolute success rates or compare OdoBot with Agent-E on success. The results support application-specific behavior modeling as a way to improve the efficiency of web-agent execution, while the evidence described is limited to the tested Canvas LMS tasks.