Debate mode facilitates structured discussions between AI models, allowing them to present arguments, respond to each other, and refine their positions over multiple rounds.
Usage
TUI (Interactive)
/debate <agents> "<topic>"
/settings # configure rounds & moderator
CLI
puzldai debate "<topic>" -a <agents> [-r rounds] [-m moderator]
| Flag | Description |
|---|
-a | Comma-separated agents to debate (required) |
-r | Number of debate rounds (default: 2) |
-m | Agent to synthesize the conclusion |
Examples
# TUI: Debate between Claude and Gemini
/debate claude,gemini "GraphQL vs REST for our API"
# CLI: Three-model debate with 3 rounds and moderator
puzldai debate "Monolith vs microservices" -a claude,gemini,codex -r 3 -m ollama
How It Works
Initial Positions
Each agent presents their initial position on the topic. All agents respond in parallel.
Response Rounds
In each subsequent round, agents read all previous responses and can:
- Defend their position with new arguments
- Address counterpoints from other agents
- Update their view if convinced
Conclusion (Optional)
A moderator agent can synthesize the debate into a balanced conclusion, summarizing agreements and disagreements.
Configuration
TUI Settings
Configure debate options via /settings → Collaboration tab:
| Setting | Description |
|---|
| Rounds | Number of debate rounds (1-5, default: 2) |
| Moderator | Agent to synthesize conclusion (default: none) |
CLI Flags
# 3 rounds with ollama as moderator
puzldai debate "GraphQL vs REST" -a claude,gemini -r 3 -m ollama
When to Use Debate
Debate mode excels at:
- Exploring trade-offs - Different perspectives highlight pros and cons
- Technical decisions - Architecture, technology choices, design patterns
- Edge case discovery - Adversarial discussion uncovers potential issues
- Learning - See how different models reason about problems
Keyboard Shortcuts
| Key | Action |
|---|
Escape | Hide view (keeps running) |
Ctrl+E | Return to debate view |
Ctrl+C | Cancel debate |
Enter | Select winner (after completion) |
Example Output
─── Debate: GraphQL vs REST ───
Round 0 - Initial Positions:
claude: "GraphQL offers significant advantages for our use case..."
gemini: "REST remains the better choice because..."
Round 1 - Responses:
claude: "Addressing Gemini's point about caching..."
gemini: "While Claude makes valid points about flexibility..."
─── Moderator Conclusion ───
Both approaches have merit. For this specific project...
Tips
Choose complementary models - Different models often have different strengths and perspectives.
2-3 rounds is usually enough - More rounds can lead to repetition without new insights.
