Agents on a LEASH: A Case Study in Micro-Managing Web Agent Behavior

Abstract

AI web agents that perform tasks on behalf of users will transform business through massive productivity gains via automation of specific tasks using a web browser. However, current commercial agents and agentic frameworks offer few guarantees that erroneous behavior will be limited. In this paper, we present LEASH, a framework that borrows from planning and focuses on defining micro-tasks for an agent that consists of both sensing and actions. By pre-sensing before an action and post-sensing after, the agent can drastically increase the likelihood of accurate behavior and mitigate risk. We present a real use-case where a LEASH agent performs autonomous pre-screening tasks of articles submitted to the Journal of AI Research (JAIR). The LEASH agent performs the tasks accurately, at scale, and without harm. We compare this to a commercially available agent which may perform well, at times, but which also exhibited grossly uncontrolled behavior.