When an LLM-powered multi-agent system fails, developers face the daunting task of identifying which agent caused the failure and at what stage. This process, often described as 'finding a needle in a haystack,' is both time-consuming and error-prone. To address this challenge, researchers from Penn State University, Duke University, and several leading institutions introduced the concept of automated failure attribution, along with the first benchmark dataset, Who&When. In the following Q&A, we break down the key findings, methods, and implications of this research, which was accepted as a Spotlight presentation at ICML 2025.

1. What is the core problem addressed by this research?

LLM Multi-Agent systems collaborate to solve complex tasks, but failures are common—often due to a single agent's mistake, miscommunication, or information loss. Currently, developers must manually sift through extensive interaction logs to identify the root cause, a process akin to archaeological digging. This research defines the novel problem of automated failure attribution: given a failed task and the agents' conversation logs, automatically pinpoint which agent caused the failure and when (at which interaction step) it happened. This is crucial for efficient debugging and system improvement. Without such automation, tuning multi-agent systems remains laborious and relies heavily on expert intuition. The authors formally propose this task and build the first benchmark to evaluate solutions, aiming to replace manual log archaeology with reliable, scalable attribution methods.

2. Why is debugging multi-agent systems so difficult?

Debugging these systems presents unique challenges. First, manual log archaeology requires developers to read long, interleaved conversations among agents, tracing the flow of information—a task that quickly becomes overwhelming as the number of agents or steps grows. Second, the process demands deep expertise: understanding each agent's role, prompt, and behavior. Misdiagnosis is common because a failure may stem from a subtle misunderstanding rather than an obvious error. Third, the autonomous nature of agent collaboration means interactions are unpredictable, making it hard to replay or simulate failures. Finally, information chains can be long: a mistake early in the pipeline propagates and corrupts later steps. These factors combine to make debugging a multi-agent system much harder than debugging a single LLM call, motivating the need for automated attribution tools that can quickly isolate the responsible agent and interaction step.

3. What is the Who&When benchmark dataset?

Who&When is the first benchmark specifically designed for automated failure attribution in LLM multi-agent systems. It contains hundreds of task instances where agents collaboratively work on tasks (e.g., planning, reasoning, code generation) and where some tasks fail. For each failure, the dataset provides ground-truth labels indicating which agent caused the failure and at which specific interaction step. The benchmark covers diverse failure types, including errors from individual agents, miscommunication between agents, and cascading failures. By releasing Who&When on Hugging Face, the researchers enable the community to develop and compare attribution methods on a standardized test bed. The dataset also includes full interaction logs, allowing models to exploit conversational context. This resource is a foundational step toward making multi-agent systems more reliable and debuggable.

4. What automated attribution methods were developed or evaluated?

The researchers explored several automated approaches. Direct prompting uses a powerful LLM (e.g., GPT‑4 or Claude) to analyze the logs and answer: 'Which agent failed?' along with the failure step. Step-wise scanning checks each agent's output at each step for anomalies or contradictions using a separate evaluator model. Causal influence methods simulate what would happen if an agent's output were replaced or modified, measuring the impact on the final result. Graph-based attribution models the information flow and computes error propagation probabilities. All methods were evaluated on the Who&When benchmark under different failure scenarios. Results show that while direct prompting works reasonably, step-wise scanning with a dedicated evaluator achieves higher accuracy, especially for subtle failures. However, no single method dominates all cases, indicating room for improvement. The full evaluation details are available in the paper.

5. How was this research received and what resources are available?

The paper was accepted as a Spotlight presentation at ICML 2025, one of the top machine learning conferences, signaling the community's recognition of the problem's importance. To support further research, the authors have fully open-sourced both the code and the Who&When dataset. This allows anyone to reproduce the experiments, develop new attribution methods, and extend the benchmark. The release includes scripts for generating synthetic failures, baseline implementations, and evaluation metrics. By making these resources openly available, the team hopes to accelerate progress in debugging and reliability of multi-agent systems, eventually leading to more robust LLM-based applications.

6. Which institutions and authors are behind this work?

The research is a collaborative effort involving multiple leading institutions: Penn State University and Duke University (co-first authors Shaokun Zhang and Ming Yin), along with Google DeepMind, University of Washington, Meta, Nanyang Technological University, and Oregon State University. This diverse team brings together expertise in LLM agents, software debugging, and benchmarking. The co-first authors are Shaokun Zhang (Penn State) and Ming Yin (Duke). The combined experience of the team ensured a thorough analysis of the problem and robust experimental design. Their collaboration highlights the interdisciplinary nature of emerging challenges in AI system reliability and diagnostics.

7. What future directions does this work open up?

This research lays the foundation for automated failure attribution, but many avenues remain. Improved attribution methods could integrate causal reasoning, attention analysis, or even adversarial probes to better isolate failure sources. Real-time attribution would be valuable for systems that need to self-correct during task execution. Extensions to heterogeneous agents (different LLMs or tool-use) are needed. The Who&When benchmark can also be expanded to include more task types, failure modes, and multi-step chains. Ultimately, automated attribution could be integrated into development pipelines, enabling faster iteration and more reliable multi-agent applications. This work signals a shift from building agents to understanding and repairing them—a critical step for deploying LLM systems in high-stakes domains.