ptrnsai

Blind Delegation

Intermediate🚫 Anti-Pattern💥 Anti-Patterns: WorkflowIndustry observation
🚫Anti-Pattern— This describes a common mistake to avoid, not a pattern to follow.

The Anti-Pattern

A supervisor agent delegates to sub-agents without verifying their output, trusting results uncritically.

Why It Happens

The supervisor or orchestrator assumes workers always produce correct output. No validation, no sanity checks, no quality gates at delegation boundaries. One bad worker output cascades through the entire system. The supervisor degrades from a quality controller to a dumb router that adds cost without adding reliability.

How to Fix It

Implement output validation at every delegation boundary. Use evaluator-optimizer or consensus voting for critical subtasks. The supervisor’s job isn’t just to route — it’s to verify. Every delegation should have a validation step that checks the worker’s output before passing it downstream. The cheaper alternative: structured output schemas that make invalid results syntactically impossible.

Diagram

  Blind (cascading failure):          Verified (caught at boundary):
  ┌────────┐    ┌────────┐           ┌────────┐    ┌────────┐
  │ Super  │───▶│Worker A│           │ Super  │───▶│Worker A│
  └────────┘    └───┬────┘           └────────┘    └───┬────┘
                    │ (bad output)                     │ (bad output)
                    ▼                                  ▼
               ┌────────┐                         ┌────────┐
               │Worker B│  ← builds on bad data   │Validate│──▶ ✗ Reject
               └───┬────┘                         └────────┘
                   │ (worse output)                    │ retry
                   ▼                                   ▼
              ┌─────────┐                         ┌────────┐
              │ Output  │  ← garbage              │Worker A│ (2nd attempt)
              └─────────┘                         └────────┘

Symptoms

  • Downstream agents process garbage input from upstream without complaint
  • No validation exists between agent handoff points
  • Errors are only discovered at the final output — never at intermediate steps
  • End-to-end quality is worse than a single agent doing everything

False Positives

  • Low-stakes delegation where the cost of verification exceeds the cost of errors
  • Well-tested deterministic sub-agents with known reliability characteristics
  • Pipelines where each step has its own independent error handling

Warning Signs & Consequences

Warning Signs

  • Final output quality worse than what a single agent produces
  • Errors traceable to intermediate steps that were never validated
  • No logging or monitoring at delegation boundaries
  • Sub-agents receiving nonsensical inputs and dutifully processing them

Consequences

  • Cascading errors that amplify through each delegation step
  • Debugging nightmares — the root cause is 3 agents upstream from the symptom
  • Lower overall quality than just using a single agent end-to-end
  • Wasted compute processing and building on invalid intermediate results

Remediation Steps

  1. 1Add output validation schemas at every delegation boundary
  2. 2Implement retry logic when worker output fails validation
  3. 3Use evaluator-optimizer pattern for high-stakes subtasks
  4. 4Log all inputs and outputs at delegation boundaries for debugging
  5. 5Set quality thresholds — reject and retry rather than pass through bad output

Real-World Example

Research Report Hallucination

A research agent delegates fact-gathering to three sub-agents, then synthesizes their findings into a report. One sub-agent hallucinates a statistic — ‘Market grew 340% in Q3’ — which the supervisor passes through unchecked. The final report prominently features this fabricated number. A simple validation step (check that cited statistics have source URLs) would have caught it.

References