Overview
In multi-agent systems, agents develop dependencies on each other's outputs. When one agent fails—through hallucination, incorrect tool use, or reasoning error—downstream agents treat that flawed output as reliable input, propagating and often amplifying the error.
Mechanism
Agent A: Makes small error (95% correct)
↓
Agent B: Builds on error (90% of A's accuracy = 85.5%)
↓
Agent C: Compounds further (90% of B's = 77%)
↓
Agent D: Now severely degraded (69%)
A 5% error rate cascades to 31% error rate across four agents.
Real-World Manifestations
Research Pipeline Degradation
A research agent misidentifies a source. The synthesis agent incorporates the error. The editing agent polishes the incorrect content. The review agent, seeing polished prose, approves it.
Financial Cascade
A data extraction agent misreads a decimal point. The calculation agent produces incorrect totals. The reporting agent presents wrong figures to stakeholders.
Agentic Workflow Corruption
In OWASP's ASI08 analysis, agentic AI cascading failures are particularly dangerous because:
- Semantic opacity: Natural language errors pass validation checks
- Emergent behavior: Multiple agents create unintended outcomes
- Temporal compounding: Errors persist in memory and contaminate future operations
Detection Signals
- Confidence scores decrease through the pipeline
- Output quality metrics diverge from expectations
- Inconsistencies appear in final outputs that weren't in inputs
- Agents produce responses outside their normal patterns
Contributing Factors
Lack of Independent Verification
Agents trust upstream outputs without cross-checking original sources.
Context Window Limitations
Downstream agents may lose context about uncertainty or caveats.
Over-Optimization for Speed
Skipping verification steps to reduce latency.
Monoculture Risk
When all agents use similar models, they share the same blind spots.