Where LLM Agents Fail and How They Can Learn From Failures

Where LLM Agents Fail and How They can Learn From Failures

📊 AgentErrorBench Dataset

Access our comprehensive benchmark dataset of systematically annotated failure trajectories:

🔗 Download AgentErrorBench

AgentErrorBench contains 200 expertly annotated agent failure trajectories across three environments:

• GAIA: 50 trajectories from general AI assistant tasks
• ALFWorld: 100 trajectories from embodied agent tasks
• WebShop: 50 trajectories from web navigation and shopping tasks

📖 About

Large Language Model (LLM) agents have shown remarkable capabilities in solving complex, multi-step tasks through sophisticated architectures integrating planning, memory, reflection, and tool-use modules. However, these complex systems are vulnerable to cascading failures, where a single root-cause error propagates through subsequent decisions, ultimately leading to task failure.

AgentDebug introduces a principled framework for understanding, detecting, and recovering from agent failures through three key contributions:

1. AgentErrorTaxonomy 📋: A modular classification system categorizing failure modes across memory, reflection, planning, action, and system-level operations.

2. AgentErrorBench 🎯: The first comprehensive dataset of systematically annotated failure trajectories from real-world agent rollouts in ALFWorld, GAIA, and WebShop environments.

3. AgentDebug Framework 🛠️: An intelligent debugging system that isolates root-cause failures and provides targeted corrective feedback, enabling agents to recover and iteratively improve.

🚀 Key Results

Our experiments demonstrate that AgentDebug significantly improves agent reliability:

• 24% higher all-correct accuracy compared to the strongest baseline
• 17% higher step accuracy in error detection
• Up to 26% relative improvement in task success rates through iterative recovery
• Effective across diverse environments (ALFWorld, GAIA, WebShop)

🏗️ Architecture

The AgentDebug framework consists of a two-stage analysis pipeline:

Stage 1: Fine-Grained Analysis

Performs detailed step-by-step analysis of agent trajectories to identify potential error patterns at each decision point.

Stage 2: Critical Error Detection

Identifies the critical failure point that led to task failure and provides root cause analysis with targeted feedback.

📁 Repository Structure

AgentDebug/
├── detector/              # Core detection and analysis framework
│   ├── fine_grained_analysis.py    # Stage 1: Step-level error detection
│   ├── critical_error_detection.py # Stage 2: Critical failure identification
│   └── error_definitions.py        # Comprehensive error taxonomy
├── assets/                # Project resources
│   └── logo.png          # AgentDebug logo
└── README.md

Note: The AgentErrorBench dataset is hosted on Google Drive due to its size. Download it from the link above.

🔧 Installation

git clone https://github.com/ulab-uiuc/AgentDebug.git
cd AgentDebug
pip install -r requirements.txt

💡 Quick Start

from detector import fine_grained_analysis, critical_error_detection

# Load your agent trajectory
trajectory = load_agent_trajectory("path/to/trajectory.json")

# Stage 1: Analyze potential errors at each step
step_errors = fine_grained_analysis.analyze(trajectory)

# Stage 2: Identify critical failure point
critical_failure = critical_error_detection.detect(trajectory, step_errors)

# Get corrective feedback
feedback = critical_failure.generate_feedback()

📊 Error Taxonomy

Our comprehensive error taxonomy covers five key modules:

Module	Error Types	Examples
Memory	Hallucination, Retrieval Failure, Over-simplification	Agent forgets visited locations
Reflection	Progress Misjudgment, Outcome Misinterpretation	Incorrect assessment of task progress
Planning	Inefficient Planning, Constraint Ignorance	Selecting impossible actions
Action	Format Errors, Parameter Errors, Misalignment	Invalid action syntax
System	Step Limits, Tool Failures, Environment Errors	Exceeding maximum steps

📈 Performance

AgentDebug achieves state-of-the-art performance in error detection and recovery:

Metric	Improvement
All-Correct Accuracy	+24%
Step Accuracy	+17%
Task Success Rate	Up to +26%

📝 Citation

If you use AgentDebug in your research, please cite our paper:

@article{agentdebug2025,
  title={Where LLM Agents Fail and How They Can Learn From Failures},
  author={Zhu, Kunlun and Liu, Zijia and Li, Bingxuan and Tian, Muxin and Yang Yingxuan and Zhang, Jiaxun and others},
  journal={arXiv preprint arXiv:2509.25370},
  year={2025}
}

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🤝 Contributing

We welcome contributions! Please feel free to submit issues, create pull requests, or reach out for collaborations.