R-MC-10 Multi-Agent & Coordination DAMAGE 3.7 / High

Sigma Degradation Cascade

Upstream agent's low process sigma degrades downstream agent's effective sigma. Quality ceiling cascades through the agent chain. No individual agent is below threshold but system quality is unacceptable.

The Risk

In manufacturing and quality management, "six sigma" refers to process quality with defect rate below 3.4 per million. The concept applies to agentic workflows: each agent has a "process sigma," the quality level at which it operates. A 3-sigma agent has roughly 66K defects per million. A 6-sigma agent has roughly 3.4 defects per million.

Sigma Degradation Cascade occurs when: (1) Upstream Agent A operates at 4-sigma (63 defects per million); (2) Downstream Agent B depends on Agent A's output and operates at 5-sigma (233 defects per million); (3) Agent B's effective sigma is degraded to roughly 4-sigma because a fraction of its input is corrupted by Agent A; (4) Over N steps, system sigma degrades below acceptable thresholds.

In regulated industries, cascade degradation means that a chain of agents individually meeting quality thresholds collectively produces system quality below threshold. An institution may approve deploying each agent because each meets individual SLA, without recognizing that the system SLA is violated.

How It Materializes

A commercial bank processes commercial real estate (CRE) loan applications through a multi-step agentic workflow. Each step is intended to operate at high quality: Document-Intake Agent (4-sigma: 99.38% accurate extraction), Data-Validation Agent (5-sigma), Compliance-Screening Agent (5-sigma), Risk-Scorecard Agent (5-sigma), and Approver Agent (5-sigma).

Each agent individually operates at its stated sigma. But the system operates differently. Document-Intake produces 630 errors per million (4-sigma). Data-Validation receives 630 errors per million from upstream. Data-Validation's native error rate is 23 per million (5-sigma), but it cannot detect Intake errors; it validates against business rules. A misextracted loan amount bypasses Data-Validation. Data-Validation's effective error rate becomes 630 per million.

The corruption cascades: Compliance-Screening, Risk-Scorecard, and Approver all inherit the 630-errors-per-million upstream defect rate. System defect rate: 630 per million, roughly 3.2-sigma. Over 6 months, the bank processes 10,000 CRE loan applications. At 3.2-sigma system quality, approximately 6-7 loan applications are approved/denied incorrectly due to cascading quality degradation.

When regulators audit the workflow, they note that the portfolio defect rate is much higher than individual agent accuracy would suggest, indicating cascading failure. The bank is required to redesign the workflow to either increase Document-Intake sigma, implement human review of Intake results, or reduce dependence on perfect Intake accuracy.

DAMAGE Score Breakdown

DimensionScoreRationale
D - Detectability3Cascading degradation is observable in system-level quality metrics but may not be obvious from individual agent metrics. Requires cross-agent analysis.
A - Autonomy Sensitivity3Affects sequential autonomous agent workflows. Parallel or independently-validated agents avoid cascade.
M - Multiplicative Potential4Affects every transaction processed through the cascade. Probability and severity scale with cascade length.
A - Attack Surface2Not directly exploitable, but adversary could deliberately introduce errors in upstream agents to trigger cascade.
G - Governance Gap4Institutions typically define individual agent SLAs, not system-level SLAs. Cascade risk is often unrecognized.
E - Enterprise Impact3Affects system quality, compliance, and customer satisfaction. Portfolio defect rates may exceed acceptable thresholds.
Composite DAMAGE Score3.7High. Requires dedicated mitigation controls and monitoring.

Agent Impact Profile

How severity changes across the agent architecture spectrum.

Agent TypeImpactHow This Risk Manifests
Digital AssistantLowHuman reviews all outputs and catches upstream errors. No cascade.
Digital ApprenticeLowAgents validate inputs before processing. Upstream errors are detected and flagged.
Autonomous AgentMediumAgents process independently; cascade risk depends on architecture.
Delegating AgentMediumSingle agent invoking tools in sequence can experience cascade if tool outputs degrade with each call.
Agent Crew / PipelineCriticalSequential pipeline architecture guarantees cascade. System sigma is constrained by weakest agent in pipeline.
Agent Mesh / SwarmHighPeer-to-peer agents with interdependencies can experience cascades if agents have quality dependencies.

Regulatory Framework Mapping

FrameworkCoverageCitationWhat It AddressesWhat It Misses
NIST AI RMF 1.0PartialMEASURE 5.2, MANAGE 7.3System performance measurement and management.Cascade quality degradation in sequential systems.
MAS AIRGPartialModel Risk, Process DisciplineProcess and model governance.System-level quality requirements and cascade analysis.
OCC / SR 11-7PartialOperational RiskModel governance and control.Cascade risk in sequential systems.
ISO 42001PartialSection 8.1, 8.3Resource and information management.Quality cascade and system-level SLA requirements.

Why This Matters in Regulated Industries

Regulated institutions are required to maintain acceptable quality standards in core processes. Loan origination, claims processing, and compliance screening must meet defined quality thresholds. If the institution deploys a cascade of agents where individual quality thresholds are met but system quality is not, the institution is not meeting its regulatory obligations.

Additionally, cascade degradation creates hidden risk. The institution may believe its system is high-quality (all agents individually excellent) when system quality has degraded significantly. This is particularly dangerous in risk management: a risk scoring system with degraded sigma may systematically under-estimate or over-estimate risk.

Controls & Mitigations

Design-Time Controls

  • Define system-level quality (sigma) requirements in addition to individual agent SLAs. Calculate the required individual agent sigmas to achieve system sigma target.
  • Implement input validation at each step to detect and correct upstream errors before they cascade.
  • Reduce cascade length by designing workflows to process data in parallel where possible instead of sequentially.
  • Use the Blast Radius Calculator to model the cascade impact of upstream agent degradation.

Runtime Controls

  • Monitor system-level quality metrics, not just individual agent metrics. Track portfolio defect rate, system accuracy, or end-to-end success rate.
  • Implement cross-step quality audits. Sample outputs at each step and verify quality does not degrade beyond expected cascade.
  • Track sigma degradation per workflow. For each agent in the pipeline, measure its actual output quality conditioned on input quality.

Detection & Response

  • Conduct regular sigma analysis on agentic workflows. Model expected system sigma based on individual agent sigmas and cascade length. Compare expected to actual.
  • When system quality degrades below target, perform root cause analysis tracing back through the cascade to identify the constraining upstream agent.
  • Implement automated alerts if system sigma degrades below target thresholds.
  • Use the Kill Switch to halt workflows where cascade-induced quality degradation is detected.

Related Risks

Address This Risk in Your Institution

Sigma Degradation Cascade requires architectural controls that go beyond what existing frameworks provide. Our advisory engagements are purpose-built for banks, insurers, and financial institutions subject to prudential oversight.

Schedule a Briefing