Debugging Critical Failures in Agent Trajectories

Summary & Methodology Analysis

TrajDebug manages the complexity of long-horizon debugging by constructing a hierarchy of history views, ranging from high to low detail. This architecture balances the requirement for evidence preservation with the practical challenges of long-context management. By identifying atomic error triggers, the system finds mismatches between agent actions and references such as task instructions or environment feedback. These triggers are then clustered into error instances based on shared violation targets, allowing for a structured representation of the failure. Each instance is classified by state, tracking whether the error was resolved or resulted in terminal impact through factors like irreversible changes or budget debt. The pipeline filters these instances to prioritize Costly Resolution or Manifest Active states, which are the most likely candidates for causing a total task failure. Finally, it uses an LLM to conduct candidate-set-guided causal attribution to isolate the single critical step. This process aims to solve the problem where multiple local errors coexist, often masking the primary root cause. Despite these improvements, the paper notes that detection accuracy for critical errors currently decreases as the length of the agent trajectory grows. Furthermore, standard candidate-error methods often fail to accurately determine criticality when many local errors are present, which this research aims to mitigate through its structured filtering and causal attribution workflow. The paper does not specify precise latency or compute costs associated with these inference steps.