Summary

arXiv:2603.06847 — submitted 6 March 2026.

Empirical grounded-theory study of 385 faults sampled from 13,602 GitHub issues across 40 open-source agentic AI repos; derives taxonomies of fault types, observable symptoms, and root causes specific to tool-invocation and long-horizon task execution.

Applicability to Zeph

HIGH — Direct complement to #2203 (12-category taxonomy, arXiv:2601.16280): that paper is dataset-derived and taxonomic; this one is empirically grounded in real OSS agent repos. The symptom/root-cause linkage is directly usable for designing Zeph's error-handler dispatch logic and retry signal classification in zeph-tools.

Implementation Sketch

• Map the paper's symptom taxonomy to existing ToolError variants in zeph-tools
• Use root-cause linkage to inform retry strategy in the error taxonomy design from research(reliability): 12-category tool invocation error taxonomy for targeted retry/fallback strategies (arXiv:2601.16280) #2203
• Potential: structured error classification at the ShellExecutor and WebScrapeExecutor level

References

• https://arxiv.org/abs/2603.06847
• Related: research(reliability): 12-category tool invocation error taxonomy for targeted retry/fallback strategies (arXiv:2601.16280) #2203 (12-category taxonomy), research(reliability): AgentDebug — structured corrective feedback on tool failures to prevent context corruption (arXiv:2509.25370) #2199 (AgentDebug structured feedback)