Multi Agent Orchestrator

📋 Multi-Agent Orchestrator = Competitive Parallel Execution

Core Principle: Launch N CLI agents (Claude Code, Codex, Gemini) with identical task → Compare self-evaluations → Declare winner based on measurable success criteria. 🔴 CRITICAL: NEVER MOCK DATA! Try multiple approaches to get real data; if all fail, stop and document attempts.

Multi-Agent Workflow Structure (Continuant - TD):

graph TD
    Task[Task File] --> Claude[Claude Code Workspace]
    Task --> Codex[Codex CLI Workspace]
    Task --> Gemini[Gemini Workspace]

    Claude --> CPlan[01_plan_claude_code.md]
    Claude --> CResults[90_results_claude_code.md]
    Claude --> CArtifacts[All Artifacts]

    Codex --> DPlan[01_plan_codex.md]
    Codex --> DResults[90_results_codex.md]

    Gemini --> GPlan[01_plan_gemini.md]
    Gemini --> GResults[90_results_gemini.md]

Orchestration Process Flow (Occurrent - LR):

graph LR
    A[Agree on Folder] --> B[Create Draft Task]
    B --> C[User Edits File]
    C --> D[User Says Ready]
    D --> E[Launch Parallel Agents]
    E --> F[Monitor Plan Files]
    F --> G[Compare 90_results]
    G --> H[Declare Winner]

Ontological Rule: TD for workspace structure (what exists), LR for orchestration workflow (what happens)

Primary source: algorithms/product_div/Multi_agent_framework/00_MULTI_AGENT_ORCHESTRATOR.md Session ID: e9ce3592-bd66-4a98-b0e7-fcdd8edb5d42 by Daniel Kravtsov (2025-11-13) - v2.0.0 Release log: See SKILL_RELEASE_LOG.md for full version history

🎯 When to Use

¶1 Use multi-agent framework when:

• Task complexity >7/10
• Multiple valid implementation approaches exist
• Need competitive evaluation
• Best solution critically important

¶2 Use Task tool sub-agents when:

• Single specialized capability (gmail, notion, jira)
• Standard workflow exists
• Quick operation needed
• Complexity <5/10

📐 Setup Workflow

¶1 MANDATORY FIRST STEP: Agree on Location

Before creating anything, ask:

• "Where should I create this task folder?" (suggest 2-3 options based on task type)
• "What should the folder name be?" (format: XX_descriptive_name)

Example:

🤖: "For your task, I suggest:
   1. /client_cases/[client]/15_[task]/ (if client-specific)
   2. /algorithms/product_div/15_[task]/ (if algorithm)

   Which location? And folder name?"

👤: "Use client_cases/HP/15_customer_metrics/"

🤖: "✅ Creating task in: /client_cases/HP/15_customer_metrics/"

¶2 Create Draft Task File Immediately

After folder agreement, create quick draft - user will edit directly:

mkdir -p [agreed_path]
cd [agreed_path]

cat > 01_task_multi_agent.md << 'EOF'
## Task: [Your quick understanding]

**Success Criteria:** [DRAFT - user refines]
- [Draft criterion 1]
- [Draft criterion 2]

## Instructions for User:
1. 📝 EDIT THIS FILE - Add details, fix criteria
2. ✅ CONFIRM - Reply "Ready" when good
3. 🔄 ITERATE - Edit and reply with changes

**Current Status:** 🔄 AWAITING YOUR EDITS

## Agents Artifact Requirement
Each agent MUST create:
- `01_plan_[agent].md` - Planning with progress updates
- `90_results_[agent].md` - Results with self-evaluation
- All outputs in workspace folder (claude_code/, codex_cli/, gemini/)

**Self-Evaluation Format:**
### Criterion 1: [from task]
**Status:** ✅/❌/⚠️ | **Evidence:** [data] | **Details:** [how tested]

## Overall: X/Y criteria met | Grade: ✅/❌/⚠️
EOF

mkdir -p claude_code codex_cli gemini
cd ..

echo "📄 Task file: [agreed_path]/01_task_multi_agent.md"
echo "🔗 file://[full_path]"

¶3 User Edits Task File

User has full control - edits file in IDE. No chat back-and-forth!

User workflow:

1. Open file (link provided)
2. Edit directly - improve description, refine criteria
3. Reply "Ready" or "Change criterion #2 to: [text]"

¶4 Wait for Confirmation

DO NOT PROCEED until user says "Ready".

Acceptable:

• ✅ "Ready"
• ✅ "Ready with changes: [edits]"
• ✅ "Change criterion #2 to: [text]"

🔄 Execution

¶1 Launch Parallel Agents

When user says "Ready":

# Run in background
./run_parallel_agents.sh [agreed_path]/01_task_multi_agent.md &
SCRIPT_PID=$!

# Monitor progress
ps aux | grep $SCRIPT_PID
tail -f [task_folder]/*/claude_output.log

Script location: algorithms/product_div/Multi_agent_framework/run_parallel_agents.sh

Scripts handle automatically:

• ✅ Repository root execution
• ✅ .env file loading
• ✅ Workspace setup/cleanup
• ✅ Background process management
• ✅ Real-time monitoring (updates every 5s)

Timing:

• Codex: 2-3 min
• Claude: 5+ min
• Gemini: 3-5 min

¶2 Monitor via Plan Files

Track progress:

cat [agreed_path]/claude_code/01_*_plan_claude_code.md
cat [agreed_path]/codex_cli/01_*_plan_codex.md
cat [agreed_path]/gemini/01_*_plan_gemini.md

¶3 Artifact Placement (CRITICAL)

🔴 ALL ARTIFACTS MUST BE IN AGENT WORKSPACE FOLDER

Every agent MUST create ALL outputs in assigned workspace - NEVER in external directories.

❌ WRONG:

[task]/
├── claude_code/
│   ├── 01_plan.md ✅
│   └── 90_results.md ✅
├── data_processed/
│   └── output.csv ❌ WRONG!
└── results.json ❌ WRONG!

✅ CORRECT:

[task]/
├── claude_code/
│   ├── 01_plan.md ✅
│   ├── 90_results.md ✅
│   ├── output.csv ✅
│   ├── results.json ✅
│   └── script.py ✅
└── codex_cli/
    ├── 01_plan.md ✅
    └── 90_results.md ✅

Why:

1. Traceability - know which agent created what
2. Comparison - side-by-side outputs
3. Cleanup - delete failed results cleanly
4. Reproducibility - exact inputs/outputs

¶4 Compare Self-Evaluations

No manual testing - compare 90_results_*.md files only:

┌───────────────────┬─────────┬───────┬────────┐
│ Success Criteria  │ Claude  │ Codex │ Gemini │
├───────────────────┼─────────┼───────┼────────┤
│ Process <5s       │ ❌ 6.2s │ ✅ 3.8│ ✅ 4.1 │
│ Handle bad data   │ ✅      │ ✅    │ ✅     │
│ Unique approach   │ ❌      │ ✅    │ ✅     │
├───────────────────┼─────────┼───────┼────────┤
│ CRITERIA MET      │ 1/3     │ 3/3   │ 3/3    │
└───────────────────┴─────────┴───────┴────────┘
🏆 WINNER: Tie Codex/Gemini

Winner = highest score (most ✅ criteria).

🔗 Scripts & References

¶1 Ready-to-use scripts:

Main (recommended):

./run_parallel_agents.sh task_file.md

Individual:

./run_claude_agent.sh task_file.md
./run_codex_agent.sh task_file.md
./run_gemini_agent.sh task_file.md

¶2 Bundled resources:

Scripts:

• scripts/create_task_file.sh - Generate standardized task files

References:

• references/script_usage.md - Detailed script documentation
• references/task_templates.md - Pre-built templates for common scenarios
• algorithms/product_div/Multi_agent_framework/00_MULTI_AGENT_ORCHESTRATOR.md - Full guide

When to load:

• Script errors → Load script_usage.md
• Task templates → Load task_templates.md
• Comprehensive understanding → Load 00_MULTI_AGENT_ORCHESTRATOR.md

❌ Anti-Patterns

¶1 Common mistakes:

❌ Using for simple tasks (just do directly) ❌ No clear success criteria (vague goals → vague results) ❌ Mocking data (NEVER create fake data) ❌ Skipping user confirmation (always wait for "Ready") ❌ External artifacts (all outputs in workspace folders) ❌ Subjective evaluation (use measurable criteria only)

✅ Quick Reference

¶1 Complete workflow:

1. User describes complex task
2. Verify complexity >7/10
3. Agree on folder location
4. Create draft task file
5. User edits and confirms "Ready"
6. Launch ./run_parallel_agents.sh &
7. Monitor plan files
8. Compare 90_results_*.md
9. Declare winner by criteria met
10. Document results

¶2 File templates:

# 01_plan_[agent].md
## My Approach ([agent])
- [ ] Step 1: [action]
## Progress: ✅ [timestamp] Step 1 complete

# 90_results_[agent].md
## Self-Evaluation ([agent])
### Criterion 1: [from task]
**Status:** ✅/❌/⚠️ | **Evidence:** [data] | **Details:** [tested how]
## Overall: X/Y criteria | Grade: ✅/❌/⚠️

¶3 Folder structure:

[agreed_path]/
├── 01_task_multi_agent.md    # User-editable
├── claude_code/              # Claude workspace
│   ├── 01_*_plan_claude.md
│   └── 90_*_results_claude.md
├── codex_cli/                # Codex workspace
│   ├── 01_*_plan_codex.md
│   └── 90_*_results_codex.md
└── gemini/                   # Gemini workspace
    ├── 01_*_plan_gemini.md
    └── 90_*_results_gemini.md

Meta Note: See knowledge-framework skill for MECE/BFO principles. Multi-agent orchestrator uses CLI agents (not sub-agents), requires measurable success criteria, and selects winner through objective self-evaluation comparison.