A web-agent architecture called OdoBot reduces repeated interface analysis by learning a behavioral model of the application from successful demonstrations. It then uses that model to plan and execute new tasks described in natural language.
Conventional web agents repeatedly inspect a page’s interface and send large representations to a language model. OdoBot instead captures how the underlying application behaves, allowing the agent to reuse knowledge about available actions and transitions rather than rediscovering them in every run.
The researchers evaluated 45 tasks in the Canvas learning-management system. OdoBot consumed 44% fewer tokens than Agent-E and 80% fewer than WebVoyager. It also achieved a higher task-success rate than WebVoyager, though the abstract does not give the exact success figures or say that it beat Agent-E on completion.
The result suggests application-specific learning can lower the operating cost of repetitive automation. Its scope is narrow: one web application, 45 tasks and demonstrations of successful behavior. Interfaces also change, requiring the behavioral model to be updated. The work is a new arXiv preprint, so broader tests are needed before assuming the savings carry over to unfamiliar sites or less structured workflows.