multi-agent-parallel-verification

name: multi-agent-parallel-verification description: Creates parallel agent systems where one agent performs tasks while another fact-checks and verifies the work, preventing hallucinations and errors through adversarial collaboration. Use when research tasks or content generation require accuracy validation. Trigger: "fact-check", "verify research", "parallel agents", "adversarial validation". source: Transcript File Import + https://www.youtube.com/watch?v=DEMO generated_by: SkillJack (skilljacked.com)

Multi-Agent Parallel Verification

Implement parallel agent verification where one agent performs tasks while another critically analyzes and fact-checks the output in real-time.

Quick Start

Set up two agents: Agent A performs research/task, Agent B fact-checks Agent A's output. Both communicate to ensure accuracy before final delivery.

Core Workflow

When research or content generation keeps producing errors:

1. Split into parallel roles

   • Primary Agent: Performs the main task (research, writing, analysis)
   • Verification Agent: Fact-checks, validates sources, identifies inconsistencies

2. Establish adversarial communication

   • Primary agent shares findings with verification agent
   • Verification agent challenges claims and checks sources
   • Both agents iterate until consensus reached

3. Configure parallel execution

   • ALWAYS run both agents simultaneously when possible
   • Let Claude detect parallelization opportunities automatically
   • Accept increased token usage for accuracy gains

Task Distribution Pattern:

Research Agent: "Find capabilities comparison between X and Y"
Fact-Check Agent: "Verify all claims from Research Agent using provided sources"
Communication: Both agents share findings and resolve discrepancies

Techniques

Adversarial Setup

• Role Opposition: One agent advocates, other challenges
• Source Cross-Validation: Verification agent checks all citations
• Real-time Feedback: Agents communicate during task execution, not just at end

Communication Protocol

• Primary agent presents findings with sources
• Verification agent identifies potential errors or gaps
• Both agents discuss discrepancies before finalizing
• Final output requires agreement from both agents

Anti-Patterns

NEVER rely on single agent for accuracy-critical tasks

• Single agents hallucinate facts even with sources provided
• Manual correction becomes repetitive and defeats automation purpose

NEVER ignore token cost vs accuracy trade-off

• Parallelization increases token usage significantly
• But manual correction effort often exceeds token costs

NEVER skip source validation

• Verification agent must check actual sources, not just claims
• Both agents must reference same source material

Edge Cases & Error Handling

When agents disagree persistently:

• Require both agents to cite specific sources for disputed claims
• Escalate to human review for final decision
• Document disagreement points for future reference

When parallelization fails:

• Fall back to sequential verification
• Primary agent completes task, then verification agent reviews
• Maintain adversarial relationship even in sequential mode

Source availability issues:

• Verification agent flags when sources are inaccessible
• Both agents work with available materials and document limitations
• Primary agent provides alternative sources when possible

Bundled Resources Plan

• templates/agent-roles.md - Role definitions for primary and verification agents
• prompts/adversarial-setup.txt - Communication protocols between agents
• scripts/parallel-task-config.py - Configuration for parallel agent execution