Tutorial 12: Multi-Agent Debate

Time: 60 minutes | Difficulty: Advanced

Multi-agent debate systems leverage diverse perspectives and critical thinking by having multiple agents discuss, challenge, and refine ideas. This pattern produces more robust, well-reasoned outputs through collaborative argumentation.

🎯 Learning Objectives

By the end of this tutorial, you'll be able to:

Implement multi-agent debate protocols
Design agents with different perspectives
Build consensus mechanisms
Handle disagreements and conflicts
Synthesize insights from debates
Apply debate to decision-making
Understand when debate improves outcomes

🏗️ What We're Building

A debate system featuring:

Multiple Agents - Different viewpoints and roles
Debate Protocol - Structured argumentation rounds
Challenger Agents - Devil's advocate and critics
Moderator - Manages flow and synthesizes
Consensus Builder - Finds common ground
Decision Framework - Converts debate to action

📋 Prerequisites

Make sure you have:

Completed Tutorial 11: Hierarchical Agents
Understanding of argumentation and logic
PHP 8.1+ installed
Claude PHP SDK configured

🤔 What is Multi-Agent Debate?

Multi-agent debate involves multiple AI agents with different roles or perspectives discussing a topic to reach better conclusions.

Debate vs Single Agent

Single Agent:

Question: Should we adopt technology X?
Agent: Yes, because... [one perspective]
Done.

Multi-Agent Debate:

Question: Should we adopt technology X?

Proponent: Yes, because benefits A, B, C...
Opponent: No, because risks X, Y, Z...
Analyst: Data shows...
Critic: Both sides overlook...
Moderator: Considering all views...

Result: Nuanced, well-reasoned decision

🔑 Key Concepts

1. Agent Roles

Different agents bring different perspectives:

$roles = [
    'proponent' => [
        'perspective' => 'Support the proposal',
        'system' => 'You advocate for the proposal. Find benefits and opportunities.'
    ],
    'opponent' => [
        'perspective' => 'Challenge the proposal',
        'system' => 'You oppose the proposal. Identify risks and drawbacks.'
    ],
    'analyst' => [
        'perspective' => 'Objective analysis',
        'system' => 'You analyze facts objectively. Focus on data and evidence.'
    ],
    'critic' => [
        'perspective' => 'Critical thinking',
        'system' => 'You identify logical flaws and assumptions in arguments.'
    ],
    'moderator' => [
        'perspective' => 'Synthesis and balance',
        'system' => 'You synthesize viewpoints and find balanced conclusions.'
    ]
];

2. Debate Protocol

Structured rounds ensure comprehensive discussion:

$protocol = [
    'round_1' => [
        'type' => 'opening_statements',
        'agents' => ['proponent', 'opponent'],
        'purpose' => 'Present initial positions'
    ],
    'round_2' => [
        'type' => 'rebuttals',
        'agents' => ['opponent', 'proponent'],
        'purpose' => 'Challenge opponent arguments'
    ],
    'round_3' => [
        'type' => 'analysis',
        'agents' => ['analyst', 'critic'],
        'purpose' => 'Objective evaluation'
    ],
    'final' => [
        'type' => 'synthesis',
        'agents' => ['moderator'],
        'purpose' => 'Unified conclusion'
    ]
];

3. Argument Quality

Evaluate argument strength:

$argumentMetrics = [
    'logic' => 'Is reasoning sound?',
    'evidence' => 'Are claims supported?',
    'completeness' => 'Are counterarguments addressed?',
    'relevance' => 'Does it address the question?'
];

4. Consensus Building

Find common ground:

function findConsensus($debateHistory) {
    $agreementPoints = [];
    $disagreementPoints = [];
    
    foreach ($debateHistory as $round) {
        // Identify where agents agree
        $commonalities = extractCommonPoints($round);
        $agreementPoints = array_merge($agreementPoints, $commonalities);
        
        // Track persistent disagreements
        $conflicts = extractConflicts($round);
        $disagreementPoints = array_merge($disagreementPoints, $conflicts);
    }
    
    return [
        'consensus' => $agreementPoints,
        'open_issues' => $disagreementPoints
    ];
}

💡 Implementation Patterns

Basic Debate System

class DebateSystem {
    private $client;
    private $agents = [];
    private $history = [];
    
    public function __construct($client) {
        $this->client = $client;
    }
    
    public function addAgent($name, $role, $systemPrompt) {
        $this->agents[$name] = [
            'role' => $role,
            'system' => $systemPrompt
        ];
    }
    
    public function debate($topic, $rounds = 3) {
        $context = "Topic: {$topic}\n\n";
        
        for ($round = 1; $round <= $rounds; $round++) {
            echo "Round {$round}:\n";
            
            foreach ($this->agents as $name => $config) {
                $prompt = $context . "As the {$config['role']}, provide your perspective.";
                
                $response = $this->client->messages()->create([
                    'model' => 'claude-sonnet-4-5',
                    'max_tokens' => 1024,
                    'system' => $config['system'],
                    'messages' => [['role' => 'user', 'content' => $prompt]]
                ]);
                
                $statement = extractTextContent($response);
                $this->history[] = [
                    'round' => $round,
                    'agent' => $name,
                    'statement' => $statement
                ];
                
                $context .= "\n{$name}: {$statement}\n";
                echo "{$name}: {$statement}\n\n";
            }
        }
        
        return $this->synthesize($topic);
    }
    
    private function synthesize($topic) {
        $debateText = "";
        foreach ($this->history as $entry) {
            $debateText .= "Round {$entry['round']} - {$entry['agent']}:\n";
            $debateText .= "{$entry['statement']}\n\n";
        }
        
        $prompt = "Topic: {$topic}\n\nDebate transcript:\n{$debateText}\n" .
                 "Synthesize this debate into a balanced conclusion that:\n" .
                 "1. Identifies key agreements\n" .
                 "2. Acknowledges valid concerns\n" .
                 "3. Provides actionable recommendation";
        
        $response = $this->client->messages()->create([
            'model' => 'claude-sonnet-4-5',
            'max_tokens' => 1536,
            'system' => 'You synthesize debates into clear, balanced conclusions.',
            'messages' => [['role' => 'user', 'content' => $prompt]]
        ]);
        
        return extractTextContent($response);
    }
}

Structured Debate Agent

class DebateAgent {
    private $client;
    private $name;
    private $perspective;
    private $systemPrompt;
    
    public function __construct($client, $name, $perspective, $systemPrompt) {
        $this->client = $client;
        $this->name = $name;
        $this->perspective = $perspective;
        $this->systemPrompt = $systemPrompt;
    }
    
    public function respond($topic, $context, $roundType) {
        $prompt = $this->buildPrompt($topic, $context, $roundType);
        
        $response = $this->client->messages()->create([
            'model' => 'claude-sonnet-4-5',
            'max_tokens' => 1024,
            'system' => $this->systemPrompt,
            'messages' => [['role' => 'user', 'content' => $prompt]]
        ]);
        
        return extractTextContent($response);
    }
    
    private function buildPrompt($topic, $context, $roundType) {
        $prompts = [
            'opening' => "Topic: {$topic}\n\nProvide your opening position.",
            'rebuttal' => "Topic: {$topic}\n\nPrevious arguments:\n{$context}\n\n" .
                         "Rebut the opposing arguments.",
            'analysis' => "Topic: {$topic}\n\nDebate so far:\n{$context}\n\n" .
                         "Provide objective analysis.",
            'closing' => "Topic: {$topic}\n\nFull debate:\n{$context}\n\n" .
                        "Closing statement."
        ];
        
        return $prompts[$roundType] ?? $prompts['opening'];
    }
    
    public function getName() {
        return $this->name;
    }
}

🎯 Debate Strategies

1. Pro/Con Debate

Simple two-sided debate:

$proAgent = new DebateAgent(
    $client,
    'Proponent',
    'support',
    'You advocate for the proposal. Present benefits and opportunities.'
);

$conAgent = new DebateAgent(
    $client,
    'Opponent',
    'oppose',
    'You challenge the proposal. Identify risks and drawbacks.'
);

$moderator = new DebateAgent(
    $client,
    'Moderator',
    'synthesize',
    'You create balanced conclusions from debates.'
);

2. Round-Robin Discussion

Each agent responds to all others:

foreach ($agents as $speaker) {
    $otherStatements = array_filter(
        $statements,
        fn($s) => $s['agent'] !== $speaker->getName()
    );
    
    $context = implode("\n\n", array_map(
        fn($s) => "{$s['agent']}: {$s['text']}",
        $otherStatements
    ));
    
    $response = $speaker->respond($topic, $context, 'discussion');
    $statements[] = ['agent' => $speaker->getName(), 'text' => $response];
}

3. Socratic Method

Questioner challenges assumptions:

$questioner = new DebateAgent(
    $client,
    'Questioner',
    'socratic',
    'You ask probing questions to reveal assumptions and test logic. ' .
    'Never make statements, only ask clarifying questions.'
);

$responder = new DebateAgent(
    $client,
    'Responder',
    'answer',
    'You answer questions thoughtfully and defend your position with evidence.'
);

4. Devil's Advocate

One agent challenges everything:

$devilsAdvocate = new DebateAgent(
    $client,
    "Devil's Advocate",
    'challenge',
    'You challenge every assumption. Find flaws, identify risks, ' .
    'question evidence. Be skeptical of all claims.'
);

📊 Advanced Patterns

Weighted Voting

Agents vote on final decision:

function weightedVote($agents, $topic, $options) {
    $votes = [];
    
    foreach ($agents as $agent) {
        $vote = $agent->vote($topic, $options);
        $confidence = $agent->getConfidence();
        
        $votes[$vote] = ($votes[$vote] ?? 0) + $confidence;
    }
    
    arsort($votes);
    return array_key_first($votes);
}

Iterative Refinement

Debate outcome feeds back for improvement:

function iterativeDebate($topic, $agents, $iterations = 3) {
    $proposal = initialProposal($topic);
    
    for ($i = 0; $i < $iterations; $i++) {
        $feedback = debate($proposal, $agents);
        $issues = extractIssues($feedback);
        $proposal = refine($proposal, $issues);
    }
    
    return $proposal;
}

Consensus Threshold

Debate continues until agreement reached:

function debateUntilConsensus($topic, $agents, $threshold = 0.8) {
    $round = 1;
    $maxRounds = 10;
    
    while ($round <= $maxRounds) {
        $statements = collectStatements($topic, $agents);
        $agreement = measureAgreement($statements);
        
        if ($agreement >= $threshold) {
            return synthesize($statements);
        }
        
        $round++;
    }
    
    return "Consensus not reached";
}

Dynamic Agent Addition

Add specialists as needed:

function adaptiveDebate($topic, $baseAgents) {
    $debate = initializeDebate($topic, $baseAgents);
    
    // Identify gaps in discussion
    $gaps = analyzeGaps($debate);
    
    // Add specialists for gaps
    foreach ($gaps as $gap) {
        $specialist = createSpecialist($gap);
        $debate->addAgent($specialist);
    }
    
    return $debate->continue();
}

🎨 Use Cases

1. Technical Decision Making

Topic: Choose database technology
Agents: Performance expert, Cost analyst, Operations engineer, Developer
Result: Balanced decision considering all factors

2. Product Design

Topic: New feature design
Agents: User advocate, Engineer, Designer, Business analyst
Result: Feature balancing UX, feasibility, and business value

3. Risk Assessment

Topic: Project risks
Agents: Optimist, Pessimist, Realist, Risk manager
Result: Comprehensive risk identification and mitigation

4. Ethical Analysis

Topic: AI policy decision
Agents: Ethicist, Legal expert, Technologist, Public representative
Result: Ethically sound policy

⚙️ Configuration

Debate Parameters

$config = [
    'min_rounds' => 2,
    'max_rounds' => 5,
    'consensus_threshold' => 0.75,
    'max_turn_length' => 500,  // words
    'timeout_seconds' => 300
];

Agent Diversity

Ensure varied perspectives:

$perspectives = [
    'technical' => 0.3,   // 30% technical focus
    'business' => 0.3,    // 30% business focus
    'user' => 0.2,        // 20% user focus
    'ethical' => 0.2      // 20% ethical focus
];

📈 Measuring Debate Quality

Track effectiveness:

$metrics = [
    'rounds_to_consensus' => 3,
    'unique_points_raised' => 15,
    'arguments_per_agent' => [
        'proponent' => 5,
        'opponent' => 6,
        'analyst' => 4
    ],
    'agreement_level' => 0.85,
    'decision_quality_score' => 8.5
];

✅ Checkpoint

Before moving on, make sure you understand:

Multi-agent debate protocols
Different agent roles and perspectives
Structured argumentation rounds
Consensus building mechanisms
Synthesis of diverse viewpoints
When debate improves decisions
Cost-benefit of multi-agent systems

🚀 Next Steps

You've mastered Multi-Agent Debate! But what about integrating external knowledge?

Tutorial 13: RAG Pattern →

Learn how to augment agents with retrieval for knowledge-grounded responses!

💻 Try It Yourself

Run the complete working example:

php tutorials/12-multi-agent-debate/debate_agent.php

The script demonstrates:

✅ Pro vs Con debate structure
✅ Multi-round argumentation
✅ Four-agent round table discussion
✅ Moderator synthesis
✅ Consensus building
✅ Structured debate protocols

💡 Key Takeaways

Multiple perspectives improve decisions - Diverse views catch blind spots
Structure prevents chaos - Clear protocols keep debates productive
Roles create focus - Each agent contributes uniquely
Synthesis is critical - Convert debate to actionable conclusion
Balance cost and quality - More agents = better but more expensive
Consensus takes time - Allow sufficient rounds
Not always needed - Simple decisions don't require debate
Measure effectiveness - Track if debate improves outcomes

📚 Further Reading

Research Papers

Improving Factuality via Multi-Agent Debate - Du et al., 2023
Debating with More Persuasive LLMs - Khan et al., 2024
Society of Mind - Minsky, 1986

🎓 Practice Exercises

Try multi-agent debate for:

Technology Selection - Agents with different priorities (cost, performance, maintainability)
Feature Prioritization - User advocate vs business vs engineering
Risk Analysis - Optimist vs pessimist vs realist
Design Review - Multiple design philosophies debate best approach

🔧 Troubleshooting

Issue: Agents agree too quickly

Solution: Strengthen opposing viewpoints, reward finding issues

Issue: Debate never converges

Solution: Lower consensus threshold, add moderator with tie-breaking

Issue: Repetitive arguments

Solution: Track previous points, penalize repetition

Issue: One perspective dominates

Solution: Balance turn lengths, require all agents to participate

Issue: High cost with marginal benefit

Solution: Reduce rounds, use fewer agents, simpler debate structure

FilesExpand file tree

README.md

Latest commit

History