Skip to content

Multi-Agent Debate Command

Usage

/amplihack:debate <QUESTION_OR_DECISION>

Purpose

Execute multi-agent debate pattern for complex decisions. Structured debate with multiple perspectives converges on best decision through argument and synthesis.

EXECUTION INSTRUCTIONS FOR CLAUDE

When this command is invoked, you MUST:

  1. Import the orchestrator:
import sys
from pathlib import Path
sys.path.insert(0, str(Path.cwd() / ".claude/tools/amplihack"))
from orchestration.patterns.debate import run_debate
  1. Execute the pattern:
result = run_debate(
    decision_question="{QUESTION_OR_DECISION}",
    perspectives=["security", "performance", "simplicity"],  # or custom
    rounds=3,
    working_dir=Path.cwd()
)
  1. Display results:
  2. Show synthesis and recommendation
  3. Explain confidence level (HIGH/MEDIUM/LOW)
  4. Summarize key debate points
  5. Report session_id for traceability

  6. Link to logs: .claude/runtime/logs/debate_<timestamp>/

  7. Manual fallback (if orchestrator unavailable):

  8. Read workflow: .claude/workflow/DEBATE_WORKFLOW.md
  9. Execute steps manually with TodoWrite tracking

When to Use

Use for decisions with multiple valid approaches:

  • Architectural trade-offs (microservices vs monolith)
  • Algorithm selection (quick vs accurate)
  • Security vs usability decisions
  • Performance vs maintainability choices

Cost-Benefit

  • Cost: 2-3x execution time (debate rounds + synthesis)
  • Benefit: 40-70% better decision quality
  • ROI Positive when: Decision impact > 3x implementation cost

Decision Question

Execute debate for the following decision:

{QUESTION_OR_DECISION}

Configuration

The workflow can be customized by editing .claude/workflow/DEBATE_WORKFLOW.md:

  • Agent perspectives: 3 (default), 5 (extended), custom profiles
  • Debate rounds: 2-3 (standard), 4-5 (deep analysis)
  • Convergence criteria: 100% (strong), ⅔ (majority), synthesis
  • Facilitation rules

Success Metrics

From research (PR #946):

  • Decision Quality: 40-70% improvement vs single perspective
  • Blind Spot Detection: 85%+ of overlooked concerns surfaced
  • Stakeholder Alignment: 90%+ when diverse perspectives included