Graph Debates

Graph debates extend the standard linear debate format by allowing branching when agents fundamentally disagree. Instead of forcing convergence, the system explores divergent perspectives in parallel branches, then synthesizes findings.

Overview

In a traditional debate, agents take turns in a linear sequence of rounds. Graph debates introduce a tree structure where:

• When high disagreement is detected, the debate branches into parallel tracks
• Each branch explores a different perspective independently
• Branches can be merged back using synthesis strategies
• The result is a directed acyclic graph (DAG) of debate nodes

API Endpoints

POST /api/debates/graph

Run a graph-structured debate with automatic branching.

Request Body:

{
  "task": "Should we prioritize renewable energy or nuclear power?",
  "agents": ["anthropic-api", "openai-api", "gemini"],
  "max_rounds": 5,
  "branch_policy": {
    "min_disagreement": 0.7,
    "max_branches": 3,
    "auto_merge": true,
    "merge_strategy": "synthesis"
  }
}

Parameters:

Field	Type	Default	Description
task	string	required	The debate topic or question
agents	string[]	["anthropic-api", "openai-api"]	Agent names to participate
max_rounds	int	5	Maximum rounds per branch
branch_policy	object	see below	Branching configuration

Branch Policy Options:

Field	Type	Default	Description
min_disagreement	float	0.7	Minimum semantic dissimilarity to trigger branch (0-1)
max_branches	int	3	Maximum parallel branches allowed
auto_merge	bool	true	Automatically merge branches at end
merge_strategy	string	"synthesis"	How to merge: "synthesis", "vote", "best_evidence"

Response:

{
  "debate_id": "550e8400-e29b-41d4-a716-446655440000",
  "task": "Should we prioritize renewable energy or nuclear power?",
  "graph": {
    "root_node_id": "node-001",
    "nodes": [
      {
        "id": "node-001",
        "agent": "anthropic-api",
        "round": 1,
        "branch_id": "main",
        "content": "...",
        "children": ["node-002", "node-003"]
      }
    ]
  },
  "branches": [
    {
      "id": "main",
      "parent_id": null,
      "reason": "root",
      "started_at_round": 1
    },
    {
      "id": "renewable-focus",
      "parent_id": "main",
      "reason": "disagreement_on_scalability",
      "started_at_round": 2
    }
  ],
  "merge_results": [
    {
      "merged_branches": ["main", "renewable-focus"],
      "strategy": "synthesis",
      "synthesis": "Both perspectives highlight valid concerns..."
    }
  ],
  "node_count": 15,
  "branch_count": 2
}

GET /api/debates/graph/{id}

Retrieve a previously run graph debate by ID.

GET /api/debates/graph/{id}/branches

Get all branches for a graph debate.

Response:

{
  "debate_id": "550e8400-e29b-41d4-a716-446655440000",
  "branches": [
    {
      "id": "main",
      "parent_id": null,
      "node_count": 8,
      "final_position": "Nuclear offers baseload reliability..."
    },
    {
      "id": "renewable-focus",
      "parent_id": "main",
      "node_count": 7,
      "final_position": "Renewables with storage can match baseload..."
    }
  ]
}

GET /api/debates/graph/{id}/nodes

Get all nodes in the debate graph.

Branching Strategies

Automatic Branching

The orchestrator monitors semantic similarity between agent responses. When disagreement exceeds min_disagreement, it creates a new branch:

Round 1: All agents respond (main branch)
Round 2: High disagreement detected (0.85)
         → Branch "nuclear-advocate" created
         → Branch "renewable-advocate" created
Round 3-5: Each branch continues independently
Round 6: Branches merged with synthesis

Branch Reasons

Branches are created with a recorded reason:

• disagreement_on_scalability - Agents disagree on scaling approach
• opposing_evidence - Agents cite contradicting evidence
• different_frameworks - Agents use incompatible analytical frameworks
• value_conflict - Agents prioritize different values

Merge Strategies

Synthesis (default)

A synthesis agent creates a unified position that acknowledges both branches:

"Both branches make valid points. The nuclear branch emphasizes
baseload reliability, while the renewable branch highlights
declining costs and distributed resilience. A hybrid approach
could leverage nuclear for baseload while scaling renewables..."

Vote

Agents vote on which branch has the stronger argument. Winner becomes the final position.

Best Evidence

The branch with the highest evidence quality score (from the trickster system) is selected.

Use Cases

Exploring Trade-offs

Graph debates excel when there are genuine trade-offs without a clear "right" answer:

• "Microservices vs monolith for a startup?"
• "Static typing vs dynamic typing for rapid prototyping?"
• "Privacy vs convenience in user data collection?"

Multi-stakeholder Analysis

When different stakeholders have legitimately different priorities:

• "How should we allocate the engineering budget?"
• "What features should we prioritize for the next release?"

Adversarial Red-teaming

Intentionally fork branches to explore attack vectors:

• "How could this API be abused?"
• "What are the failure modes of this architecture?"

Example: Full Graph Debate

import httpx

async def run_graph_debate():
    async with httpx.AsyncClient() as client:
        response = await client.post(
            "https://api.aragora.ai/api/debates/graph",
            json={
                "task": "Should AI models be open-source or closed?",
                "agents": ["anthropic-api", "openai-api", "gemini"],
                "max_rounds": 4,
                "branch_policy": {
                    "min_disagreement": 0.6,
                    "max_branches": 2,
                    "merge_strategy": "synthesis"
                }
            },
            headers={"Authorization": "Bearer YOUR_TOKEN"}
        )
        result = response.json()

        print(f"Debate ID: {result['debate_id']}")
        print(f"Branches created: {result['branch_count']}")
        print(f"Total nodes: {result['node_count']}")

        for merge in result.get("merge_results", []):
            print(f"\nMerge synthesis:\n{merge['synthesis']}")

Integration with Other Features

• Trickster: Each branch is monitored for hollow consensus independently
• Genesis: Evolved agents can be spawned for specific branches
• ELO Rankings: Branch winners contribute to agent rankings
• Memory: Branch explorations are stored for future reference