Neura Market
Why Do Web Agents Fail? A Hierarchical Planning Perspective (2026) logo

Why Do Web Agents Fail? A Hierarchical Planning Perspective (2026)

Free

Decomposes web agent behavior into high-level planning, low-level grounding, and replanning — PDDL-structured plans outperform NL plans but grounding remains the dominant bottleneck; a single round of exploratory replanning substantially improves task success

FreeFree tier
Type
Open Source
Visit Website GitHub (40 stars)

About Why Do Web Agents Fail? A Hierarchical Planning Perspective (2026)

This paper from ACL 2026 proposes a hierarchical planning framework to analyze failures in LLM-based web agents across three layers: high-level planning, low-level execution, and replanning. Through experiments, the authors find that structured PDDL plans produce more concise and goal-directed strategies than natural language plans, but low-level execution (perceptual grounding) remains the dominant bottleneck. A single round of exploratory replanning significantly improves task success. The work provides a principled foundation for diagnosing and advancing web agents.

Key Features

Hierarchical planning framework with three layers (high-level planning, low-level execution, replanning)
Process-based evaluation enabling detailed failure analysis
Comparison of PDDL-structured plans vs. natural language plans
Identification of low-level execution (perceptual grounding) as the dominant bottleneck
Demonstration that a single round of exploratory replanning substantially improves task success
Accepted to ACL 2026

Pros & Cons

Pros
  • Provides a structured, principled framework for analyzing web agent failures
  • Clearly identifies that low-level execution, not just high-level reasoning, is the main bottleneck
  • Offers actionable insight that replanning can improve task success
  • Empirically compares PDDL and NL planning, showing advantages of structured plans
Cons
  • Evaluation is limited to specific tasks and may not generalize across all web environments
  • Findings are derived from LLM-based agents; applicability to other agent architectures is unclear
  • Applying the framework to new agents requires manual annotation of planning stages

Best For

Diagnosing failure modes in LLM-based web agentsImproving web navigation systems by focusing on grounding and adaptive controlAcademic research on hierarchical planning and agent evaluation

FAQ

Why do LLM-based web agents fail?
The hierarchical analysis shows that while high-level planning can be improved with structured formats like PDDL, the primary bottleneck is low-level execution (perceptual grounding). Failures often stem from the agent's inability to accurately perceive and interact with web elements.
What is the hierarchical planning framework?
It decomposes web agent behavior into three layers: high-level planning (task decomposition), low-level execution (grounding actions to web elements), and replanning (adapting when execution fails). This enables process-based evaluation rather than just end-to-end success.
How does PDDL planning compare to natural language planning?
PDDL-structured plans produce more concise and goal-directed strategies than natural language plans, leading to better high-level reasoning. However, even with PDDL, low-level execution remains the dominant bottleneck.