Task-Agnostic Environment Preprocessing for LLM Agents

This research explores how LLM agents can autonomously prepare unfamiliar environments by constructing indices, scripts, or procedural guidance before task execution. Unlike existing methods that rely on task-specific trajectories, this approach allows agents to explore under a budget without knowing the downstream task distribution.