Token Efficient Task Execution via Application Behavior Modeling for Web Agents
| Source: arXiv AI
Tags: web agents, token efficiency, OdoBot, Canvas LMS, WebVoyager, Agent-E
OdoBot, a web agent that builds a behavioral model of target applications from prior task demonstrations, uses 44% fewer tokens than Agent-E and 80% fewer than WebVoyager on 45 Canvas LMS tasks—while also surpassing WebVoyager on task success rate.
Details
Most web agents analyze a web app's UI from scratch on every task, incurring high token costs. OdoBot takes a different approach: it constructs a behavioral model of the target application by analyzing past successful task-execution demonstrations, then uses that model to guide future executions with far fewer tokens.\n\nTested on 45 tasks in the Canvas Learning Management System, OdoBot uses 44% fewer tokens than Agent-E and 80% fewer tokens than WebVoyager. It also surpasses WebVoyager's task success rate, though the paper does not report the specific success rate numbers for direct comparison beyond that claim.\n\nThe behavioral modeling approach is architecturally notable: instead of building application knowledge into the model via fine-tuning, it externalizes that knowledge as a behavioral model that can be updated with new demonstrations. This has practical implications for enterprise web automation where the same application is accessed repeatedly—initial demonstration runs pay down the cost of all future executions.