Debugging LLM-based agents often requires knowing which step caused a task to fail. Existing failure attribution methods rely on expensive LLM prompting and step-level error labels, which are costly to obtain. A new paper on arXiv asks whether failed trajectories themselves can reveal the decisive misstep, without such labels.
The work proposes a method that analyzes failed trajectories to attribute failure to specific steps. By avoiding step-level labels, it aims to reduce the overhead of debugging agentic systems. The authors argue this is critical for improving reliability as agents take on more complex tasks.
While the abstract does not provide experimental details, the framing suggests a shift toward more efficient, label-free failure analysis. If successful, this could lower the barrier to diagnosing agent errors in real-world deployments.