Hierarchical Planning Agent for Web-Browsing Tasks

Recent advances in large language models (LLMs) have enabled the development of agentic systems for sequential decision-making. Such agents must perceive their environment, reason over multiple time steps, and take actions that optimize long-term goals. However, existing web agents perform poorly on complex long-horizon tasks due to key limitations: limited in-context memory for tracking history, weak planning that fails to satisfy user constraints, difficulty handling task complexity, and greedy behaviors that cause premature termination. To address these challenges, we propose Structured Agent, a hierarchical planning framework with two core components: (1) an online hierarchical planning algorithm that uses dynamic AND/OR trees for efficient search, and (2) a structured memory module that tracks candidate solutions to improve constraint satisfaction in information-seeking tasks. Our preliminary experiments on WebVoyager and custom shopping benchmarks demonstrate that Structured Agent achieves improvements in long-horizon web-browsing tasks compared to standard LLM-based agents.