Designing Benchmarks and Evaluation Protocols for Agent Failure Diagnosis and Repair Capabilities

The growing deployment of LLM-based autonomous agents has exposed a critical gap in evaluation methodology: existing benchmarks predominantly measure task success under nominal conditions, offering little insight into why agents fail or whether they can recover. This paper addresses the design of benchmarks and evaluation protocols that systematically diagnose agent failures and assess repair capabilities. We argue that meaningful robustness evaluation requires (i) controlled fault injection across architectural layers, (ii) fine-grained diagnosis localizing failures to specific components-input handling, prompting, tool use, or planning-and (iii) protocols that distinguish passive success from active recovery.

Recent work has charted this territory with frameworks combining detection tools and rootcause analysis to uncover failures across agent layers, demonstrating that trace-grounded diagnosis substantially outperforms prompt-only repair strategies. Systematic perturbation taxonomies have been proposed-distinguishing transient from persistent faults and singlepath from multi-path tasks-with metrics quantifying recovery independent of task completion. Complementary evaluation frameworks have revealed significant reliability gaps between nominal and stressed performance across baseline architectures.

Yet substantial challenges remain. Benchmark validity itself is a concern, with auditing pipelines detecting hidden flaws in widely used benchmarks, confounding model errors with benchmark artifacts. Diagnostic tools must bridge the gap between curated benchmark data and real LLM-generated traces, where feature-based detectors degrade considerably.

Furthermore, the notion of "repair" requires operational definition-whether involving prompt refinement, tool schema updates, harness modification, or self-healing through closed-loop validation. This paper synthesizes these emerging threads to propose design principles for failure diagnosis and repair evaluation: composable fault scenarios targeting specific failure modes, multi-metric frameworks separating detection, localization, and repair effectiveness; regression-aware validation preventing repair-induced degradation; and architecture-agnostic protocols accommodating diverse assessment approaches. We discuss benchmark auditing as a prerequisite for trustworthy evaluation and outline open questions regarding generalization, calibration of LLM-based judges, and the transition from benchmark performance to deployment reliability.