prompt
FreeDiagnose web agent failures by hierarchical planning analysis
FreeFree tier
Inputs: textOutputs: text
About prompt
A specialized prompt for diagnosing failures in web/GUI/computer-use agent trajectories, based on the hierarchical planning perspective from the 2026 study 'Why Do Web Agents Fail?'. It analyzes failures across three layers: high-level planning (decomposing goals into subgoals), low-level grounding (mapping subgoals to concrete UI actions), and replanning (revising plans when the environment diverges). Grounded in research showing that grounding is the dominant bottleneck and that PDDL-structured plans outperform free-text plans, the prompt guides users to identify the actual bottleneck layer so that fixes target the true cause rather than wasting effort on the wrong layer.
Key Features
Analyzes failures across three layers: high-level planning, low-level grounding, and replanning
Provides evidence-based diagnosis from full agent trajectory (goals, plans, observations, actions, tool errors)
Grounds diagnosis in research findings: grounding is the dominant bottleneck, PDDL plans outperform free-text, and one replanning attempt improves success
Helps target fixes to the actual bottleneck layer instead of wasting effort on the wrong layer
Pros & Cons
Pros
- Based on peer-reviewed research (arXiv 2026) with clear empirical findings
- Structured analysis reduces guesswork in fixing agent failures
- Distinguishes between planning and grounding failures, avoiding misguided optimizations
Cons
- Requires access to full trajectory data (goals, actions, observations, tool errors) for diagnosis
- As a prompt, its effectiveness depends on the underlying LLM's ability to reason about the trajectory
Best For
Debugging failed web/GUI agent runs in research or productionImproving agent planning and grounding accuracy in web automation systemsTeaching hierarchical planning concepts for AI agent developmentAnalyzing trajectory data from browser-use or computer-use agents
FAQ
What are the three layers of web agent failure?
High-level planning (decomposing user goal into subgoals), low-level grounding (mapping subgoals to concrete UI actions), and replanning (revising plan when environment changes).
What is the dominant bottleneck according to the research?
Grounding is the dominant bottleneck. Most failures occur when a good plan hits the wrong DOM node, tab, or screen region, not from bad planning.