Strategy Mode & Multi-Agent Workflows

Overview

Strategy Mode orchestrates three AI models working together in sophisticated multi-agent workflows. Unlike Parallel Mode where models work independently, Strategy Mode creates sequential pipelines where each model's output influences the next.

This guide covers four proven multi-agent patterns:

• Ensemble Synthesis: Models draft independently, a synthesizer combines best elements
• Model Arbitration: Models compete, a judge selects the winner
• Mixture of Agents: Sequential refinement with a composer finalizing output
• Multi-Agent Debate: Models critique each other, a critic arbitrates

Prerequisites

• Completed Guide 01 (The Editable Story) and Guide 04 (Parallel Mode)
• Active subscription with sufficient credits for 3-model execution
• Understanding of different model strengths (creative vs analytical vs balanced)
• Familiarity with prompt engineering and model temperature settings

Understanding Strategy Types

Step 1: Select Strategy Mode

1. Navigate to the Authoring page
2. Click the Execution Mode dropdown (defaults to "Single")
3. Select "Strategy" from the dropdown
4. The interface will display a validation message: "Strategy mode requires exactly 3 models"

Step 2: Choose Strategy Type

1. After selecting Strategy mode, a new "Execution Strategy" dropdown appears
2. Select your desired strategy:
   • Ensemble Synthesis: For creative content needing diverse ideas
   • Model Arbitration: For competitive evaluation and selection
   • Mixture of Agents: For sequential refinement and quality improvement
   • Multi-Agent Debate: For balanced analysis of complex topics
3. The role labels will update based on your selected strategy

Step 3: Select Three Models with Role Assignment

1. Click "Available Models" to open the model selector
2. The interface now shows role-specific dropdowns (roles vary by strategy type)
3. For Ensemble Synthesis:
   • Drafter A: Choose a creative model (e.g., Claude Sonnet)
   • Drafter B: Choose a different creative model (e.g., GPT-4)
   • Synthesizer: Choose a balanced, analytical model (e.g., Gemini Pro)
4. For Model Arbitration:
   • Candidate A: Choose model with one perspective
   • Candidate B: Choose model with different strengths
   • Judge: Choose most analytical, objective model
5. For Mixture of Agents:
   • Drafter: Choose fast, creative model for initial draft
   • Critic: Choose analytical model good at finding weaknesses
   • Composer: Choose high-quality model for final polish
6. For Multi-Agent Debate:
   • Advocate: Choose persuasive, creative model
   • Challenger: Choose critical, analytical model
   • Arbitrator: Choose balanced model for synthesis

Step 4: Configure Template and Tone

1. Select an Artifact Template appropriate for your strategy:
   • Ensemble Synthesis: Marketing, Blog, Creative Writing
   • Model Arbitration: Decision Document, Technical Specification
   • Mixture of Agents: Documentation, Research, Technical Writing
   • Multi-Agent Debate: Analysis, Strategy, Research Synthesis
2. Choose Template Tone:
   • Creative (temp: 0.9): For ensemble synthesis and debate
   • Balanced (temp: 0.7): For mixture of agents
   • Precise (temp: 0.5): For model arbitration
3. Adjust Max Output Tokens if needed (higher for synthesis/composition roles)

Step 5: Write Your Prompt

Strategy mode prompts should provide clear context for the entire workflow:

Create a product launch announcement for our new AI-powered analytics platform.

Target Audience: B2B SaaS buyers (both technical and business stakeholders)

Key Points to Cover:
- Real-time data visualization
- Predictive analytics with 95% accuracy
- Integration with 50+ data sources
- Enterprise-grade security

Please generate diverse drafts that will be synthesized into a unified message.

We need to choose between two architecture approaches for our mobile app:

Option A: Native iOS/Android (Swift/Kotlin)
Option B: Cross-platform (React Native)

Context:
- Team: 5 developers (3 web, 2 mobile)
- Timeline: 6 months to launch
- Budget: $200K
- Requirement: High performance, offline-first

Generate competing analyses that will be judged for final recommendation.

Create API documentation for our User Authentication endpoint:

POST /api/auth/login

Parameters:
- email (string, required)
- password (string, required)
- remember_me (boolean, optional)

Returns: JWT token on success, error on failure

This will go through draft → critique → final composition for maximum quality.

Analyze the investment thesis for Company X:

Bullish Arguments (Advocate):
- 300% revenue growth YoY
- Strong moat with proprietary technology
- Expanding to 10 new markets

Bearish Arguments (Challenger):
- Not yet profitable ($50M loss last quarter)
- High customer acquisition cost ($500)
- Intense competition from larger players

Generate advocate position, challenger rebuttal, and arbitrated balanced assessment.

Step 6: Execute Strategy

1. Verify all three models are selected (one for each role)
2. Confirm the Generate button is enabled
3. Click Generate to start the multi-agent workflow
4. Watch the progress indicator show sequential execution:
   • Stage 1: First model generates (Drafter A / Candidate A / Drafter / Advocate)
   • Stage 2: Second model receives Stage 1 output (Drafter B / Candidate B / Critic / Challenger)
   • Stage 3: Third model receives both previous outputs (Synthesizer / Judge / Composer / Arbitrator)
5. Total execution time: 2-5 minutes depending on model speed and output length

Step 7: Review Multi-Stage Output

Strategy mode preserves all intermediate outputs for full transparency:

1. The AI Generation Pane displays a tabbed or collapsible interface
2. Each role's output is shown separately with clear labels
3. For Ensemble Synthesis:
   • Drafter A: First independent draft
   • Drafter B: Second independent draft
   • Synthesizer: Final unified output (typically the best result)
4. For Model Arbitration:
   • Candidate A: First proposal
   • Candidate B: Competing proposal
   • Judge: Evaluation, recommendation, or hybrid solution
5. For Mixture of Agents:
   • Drafter: Initial rough draft
   • Critic: Critique with improvement suggestions
   • Composer: Polished final version (usually highest quality)
6. For Multi-Agent Debate:
   • Advocate: Initial position with arguments
   • Challenger: Counter-arguments and rebuttals
   • Arbitrator: Balanced synthesis considering both sides
7. You can copy content from ANY stage, not just the final output

Step 8: Edit and Refine

Strategy mode outputs are fully editable like any other generation:

1. Click on any section to activate the Tiptap editor
2. Use the floating toolbar for text formatting (Bold, Italic, Headings, Colors)
3. Edit, delete, or reorganize sections as needed
4. You can mix content from different stages:
   • Copy introduction from Drafter A
   • Copy body paragraphs from Synthesizer
   • Copy conclusion from Drafter B
5. For Model Arbitration, you might keep both proposals and the Judge's analysis
6. For Multi-Agent Debate, you might structure final doc as: Position → Counter-Position → Synthesis

Step 9: Regenerate Specific Sections

If one section isn't meeting your needs, regenerate it without re-running the entire workflow:

1. Hover over the section header to reveal the Regenerate icon (circular arrow)
2. Click the icon to regenerate just that section
3. The regeneration uses the same model and role as the original
4. If you want a different approach:
   • Manually edit the section with your own instructions
   • Or change the execution mode to Single, select a different model, and regenerate

Step 10: Save Version with Strategy Metadata

1. Click Save Version in the toolbar
2. Enter a descriptive version name that includes the strategy type (e.g., "Product Launch - Ensemble v1")
3. The system automatically saves:
   • All three model outputs (one for each role)
   • Strategy type and configuration
   • Model assignments for each role
   • Temperature, tone, and other parameters
4. You can load this version later to see the complete multi-agent workflow
5. Versions are saved to Google Cloud Storage with immediate availability (no delay)

Step 11: Export Strategy Results

Export your final document with optional model attribution:

1. Click Export in the toolbar
2. Choose format: PDF, DOCX, or HTML
3. Toggle "Include Model Attribution" if you want footnotes showing which model generated each section
4. For Strategy mode, attribution will show role names (e.g., "Generated by Drafter: Claude Sonnet")
5. Download the file and verify formatting

Advanced: Choosing the Right Strategy

Troubleshooting

Best Practices

Summary

Congratulations! You've mastered Strategy Mode and multi-agent workflows. You now understand:

• The four core strategy types and when to use each
• How to assign models to specialized roles (Drafter, Critic, Judge, etc.)
• Sequential execution patterns and how context flows between models
• How to review and leverage intermediate outputs
• Cost optimization strategies for multi-agent workflows
• How to export strategy results with model attribution

Strategy Mode represents the cutting edge of AI-assisted content creation, enabling sophisticated multi-agent collaboration patterns that weren't possible with single-model systems.