Compound AI System Reliability: A Failure Taxonomy and Resilience Pattern Catalog from 150 Production Incidents
Source
arxiv.org
Author
Rudrendu Kumar Paul, Sourav Nandy
Date
Why it matters
Gives a failure taxonomy (retrieval, generation, tool, orchestration, integration) and patterns like circuit breakers and quality gates with reported effect sizes, so you can design for silent degradation and cascades.
Key takeaways · AI-distilled
The authors built their taxonomy from 150 production incident reports drawn from open-source compound AI projects and anonymized enterprise deployments, yielding 23 failure modes in five categories: retrieval, generation, tool, orchestration and integration failures.
Their core argument is that compound AI systems mostly break at component boundaries rather than inside a single model: errors cascade between components, quality degrades silently past standard monitoring, and correct parts can still coordinate into wrong behavior.
In the authors' controlled fault-injection experiments, circuit breakers cut cascade propagation by 89%, output quality gates caught 73% of silent degradation before users saw it, and component isolation shrank blast radius by 64%.
The paper reports that systems using three or more patterns from its catalog cut mean time to recovery by 71% compared with unstructured monitoring baselines.
articleEngineering Reliable Coding Agents: Evaluating and Operating the System Around the ModelStephanie Jarmak
articleDiagnostic Foundation for Evaluating LLMs' Research Integrity as Co-ScientistsYash Tripathi, Silu Sharma, Sai Sidhanth Manoharan Jayanthi, Shivank Garg, Lin Li
articleFinishing the Task Is Not Enough: Evaluating Agent Resilience and Considerate Participation under Accumulating ChallengeYuanchen Bai, Zijian Ding, Angelique Taylor