Skip to content

ADR 004: Implementation of Pre/Post-Execution Guardrails and Verification Layer

  • Status: Accepted
  • Date: 2026-08-06
  • Deciders: Software Architecture Team

Context and Problem Statement

LLMs can produce non-deterministic outputs, hallucinations, or security vulnerabilities (Prompt Injection, PII leakage). In a production multi-agent system, passing raw unvalidated outputs directly from one agent to another poses severe operational and security risks.

Decision Drivers

  • Security & Compliance: Prevent PII/GDPR data leakage and mitigate prompt injection attacks.
  • Data Integrity: Guarantee that agent outputs strictly conform to required JSON schemas before proceeding to downstream workflow nodes.
  • Deterministic Confidence: Verification must be an explicit platform layer, not assumed.

Decision Outcome

Chosen Option: Implement a dedicated Guardrails & Verification Pipeline between workflow node executions.

Pipeline Stages

  1. Pre-Execution Stage:
  2. Prompt Sanitization: Detect and block prompt injection patterns.
  3. PII Masking: Mask sensitive user data (emails, credit cards, SSNs) before sending prompts to external LLM providers.
  4. Post-Execution Stage:
  5. Schema Validation: Verify output matches target JSON schema.
  6. Hallucination & Toxicity Check: Evaluate confidence score; trigger retries or fallback nodes if confidence falls below threshold.

Positives

  • Enterprise Security: Complies with strict security standards.
  • Reliable Workflows: Downstream agents receive clean, verified data.