SAFARI Scales Agentic Fault Attribution Beyond Context Limits

As autonomous agents tackle increasingly complex, multi-step tasks, their execution trajectories often grow too large for even the most extensive LLM context windows. Traditional fault diagnosis methods, which load the entire trajectory into an LLM's context, suffer from attention dilution and fail when traces exceed these limits. To overcome this, researchers developed SAFARI (Scaling long-horizon Agentic Fault AttRibution via active Investigation). SAFARI fundamentally changes the diagnostic approach by replacing linear context loading with an intelligent, tool-augmented diagnostic loop. This framework provides LLMs with a specialized toolbox to read and search specific segments of a trajectory. Crucially, SAFARI also incorporates a persistent Short-Term Memory (STM) for cross-turn reasoning, allowing the LLM to maintain context and diagnose faults even when they occur far beyond its native context window. Experiments show SAFARI significantly outperforms state-of-the-art methods, maintaining high precision even when faults are 5x beyond the model's architectural context limits, a scenario where other evaluators completely fail.