A testing framework that fuzzes multi-agent LLM pipelines to find failure cascades before production

When agents chain LLM calls, errors compound. A small hallucination in step 1 becomes a confident wrong answer by step 3. The ACM called this out as a fundamental flaw in multi-LLM architectures (179 HN points). This framework takes your agent pipeline definition, generates adversarial inputs designed to trigger cascading failures, records where each stage goes wrong, and reports the failure paths with actionable guard recommendations.