AI Research

Automated Failure Attribution: New Research Pinpoints Which Agent Breaks Multi-AI Systems

Researchers from PSU and Duke introduce automated failure attribution for LLM multi-agent systems, with a new benchmark dataset called Who&When.

Researchers from PSU, Duke, and collaborators have introduced automated failure attribution for LLM multi-agent systems, creating the Who&When benchmark dataset. Experiments show current models struggle with attribution, achieving only 53.5% agent accuracy and 14.2% step accuracy, highlighting a critical gap in multi-agent system debugging.

When a multi-agent AI system fails, developers face a daunting question: which agent was responsible, and at what step did things go wrong? Sifting through lengthy interaction logs to find the root cause is like looking for a needle in a haystack.[reference:25]

Now, researchers from Penn State University and Duke University—in collaboration with Google DeepMind, the University of Washington, Meta, Nanyang Technological University, and Oregon State University—have introduced a new research area called "Automated Failure Attribution" to solve exactly this problem.[reference:26][reference:27] Their work, "Which Agent Causes Task Failures and When? On Automated Failure Attribution of LLM Multi-Agent Systems," has been accepted as a Spotlight presentation at ICML 2025.[reference:28]

The Who&When Benchmark

The team built the first benchmark dataset for this task, named Who&When. It contains failure logs from 127 diverse LLM multi-agent systems, covering complex tasks from benchmarks like GAIA and AssistantBench.[reference:29][reference:30] Each failure log is annotated with fine-grained human labels identifying which agent caused the failure, at which step, and why.[reference:31][reference:32]

The dataset includes two subsets: an algorithmically generated set with 191 unique agents across 126 logs, and a hand-crafted set with 5 unique agents across 58 logs.[reference:33]

Three Automated Attribution Methods

The researchers proposed and evaluated three automated methods for failure attribution[reference:34]:

  • All-at-Once: A single-pass LLM analysis of the entire log
  • Step-by-Step: Sequential review of each step in the conversation
  • Binary Search: A divide-and-conquer approach that searches logs efficiently

Experiments using models like GPT-4o showed the best agent-level accuracy reached only 53.5%, while step-level accuracy was a modest 14.2%—worse than random in some cases.[reference:35] Even advanced models like OpenAI o1 and DeepSeek R1 struggled, particularly with longer contexts.[reference:36] Hybrid methods improved results but increased costs.[reference:37]

Why Attribution Is So Hard

The research identifies two core challenges[reference:38]:

  • Distinguishing symptoms from root causes in multi-agent error propagation—a failure at step 10 might actually be caused by a mistake at step 3
  • Tracing information dependencies beyond temporal order—agents don't just act in sequence; they share information, and errors can propagate across the system in complex ways

The team introduced Information Dependency Graphs (IDGs) to trace causal relationships beyond temporal sequences, providing a new framework for understanding how errors spread through multi-agent systems.[reference:39]

Implications for Production AI

The findings highlight a critical gap in current AI infrastructure. Multi-agent systems are increasingly deployed in production, yet we lack reliable tools to debug them when they fail.[reference:40] The study shows that even the most advanced LLMs cannot reliably perform failure attribution today—a problem that will only grow as multi-agent systems become more complex and autonomous.

All code and data from the study have been fully open-sourced, including the dataset on Hugging Face and the code on GitHub, paving the way for further research and practical tooling.[reference:41][reference:42]

For engineering teams building agentic systems, this research represents an essential first step toward making multi-agent AI reliably debuggable—and ultimately, more trustworthy in production.