When an AI agent fails, we usually look at the model first. Maybe it is not capable enough, the prompt is unclear, or it needs better tools. But the environment in which the agent learns also matters.

An agent learns to browse a website, edit a spreadsheet, or repair code by taking actions and seeing the results. Building these environments is expensive because developers must create tasks, interaction rules, and a verifier that judges the work.

A team from Washington University in St. Louis, Google Cloud, and the University of North Carolina proposes a different approach. Its new paper, EnvHarness: Awakening Static Worlds for Agent Learning, describes a system that adapts existing environments to an agent’s weaknesses without rebuilding them.