Skip to main content

Debate Mode

Send each request to 3 different model families in parallel, then synthesize the best answer with an arbitrator model.

Debate mode is designed for high-stakes decisions where accuracy matters more than cost. It leverages diverse reasoning from multiple model families to produce a single, high-quality response.

  • 3 panelist models from distinct families (e.g. GPT, Claude, Gemini) run in parallel
  • 1 arbitrator model evaluates all responses and synthesizes the best answer
  • Automatic family selection — panelists are chosen to ensure genuinely different reasoning approaches

~4× cost per request

Debate mode runs 3 panelist models + 1 arbitrator per request. Best suited for critical queries where accuracy outweighs cost considerations. Streaming is not supported — responses are returned as a complete message.

How to Activate Debate Mode

There are three ways to enable debate mode, depending on your use case:

1. Per-request flag (API / SDK)
Add debate: true to your request body to enable debate for a single request.
javascript
import Agentlify from 'agentlify';

const client = new Agentlify({ apiKey: 'your-api-key' });

const response = await client.chat.completions.create({
  model: 'router/your-router-id',
  messages: [
    { role: 'user', content: 'Analyze the trade-offs of microservices vs monolith...' }
  ],
  debate: true  // Enable debate for this request
});

console.log(response.choices[0].message.content);
// Metadata includes debate details:
// response.metadata.debateMetadata.panelists
// response.metadata.debateMetadata.arbitratorModel
1b. Panel flag — see all responses unfiltered
Add panel: true to skip the arbitrator entirely. All 3 panelist responses are returned in a single message, each labelled with its model name, separated by horizontal rules.
javascript
const response = await client.chat.completions.create({
  model: 'router/your-router-id',
  messages: [
    { role: 'user', content: 'Explain the CAP theorem...' }
  ],
  panel: true  // Returns all 3 responses in one message, no arbitration
});

// response.choices[0].message.content will look like:
//
// **gpt-5**
//
// The CAP theorem states...
//
// ---
//
// **claude-sonnet-4-5**
//
// CAP stands for Consistency, Availability...
//
// ---
//
// **gemini-2-flash**
//
// In distributed systems...

Panel mode costs ~3× per request (no arbitrator). Useful when you want to compare model perspectives directly, or feed all responses into your own downstream logic. The response is standard OpenAI-compatible — a single choices[0].message.content string so existing SDK code requires no changes.

2. Global toggle on router
Enable debate mode for all requests to a specific router via the dashboard.
  1. Go to Dashboard → Routers
  2. Select your router and click Edit
  3. Find the Debate Mode section and toggle it on
  4. Save changes — all requests to this router will now use debate mode

Users can opt out per-request by sending debate: false in their request.

3. Preset router endpoint
Use the built-in /debate preset which comes pre-configured with debate mode enabled.
javascript
const response = await client.chat.completions.create({
  model: 'router/debate',  // Use the debate preset
  messages: [
    { role: 'user', content: 'Your question here...' }
  ]
});

How It Works

1

Panelist Selection

3 models are selected from your router's fallback list, ensuring each comes from a distinct model family. Family detection strips variant suffixes (e.g. gpt-5 and gpt-5-mini are the same family).

2

Parallel Execution

All 3 panelists receive the same request and run in parallel. The system requires at least 2 successful responses to proceed.

3

Arbitration

An arbitrator model receives all panelist responses and synthesizes the best possible answer, combining the strongest insights from each.

4

Response

The arbitrator's response is returned in the standard OpenAI-compatible format. Metadata includes details about which models were used as panelists and as the arbitrator.

Response Shape

Both modes return a standard OpenAI-compatible response. The _meta field carries mode-specific detail about cost, tokens, and which models ran.

debate: true — arbitrated response
model is the arbitrator's ID. _meta.debate lists all panelists and per-model cost.
json
{
  "id": "chatcmpl-...",
  "object": "chat.completion",
  "model": "anthropic:claude-opus-4-5",   // arbitrator
  "choices": [{ "message": { "role": "assistant", "content": "..." } }],
  "usage": { "total_tokens": 4821 },
  "_meta": {
    "modelUsed": "anthropic:claude-opus-4-5",
    "routerMode": "debate",
    "cost": 0.034210,
    "tokensUsed": 4821,
    "latency": 8342,
    "debate": {
      "arbitrator": "anthropic:claude-opus-4-5",
      "panelists": [
        { "model": "openai:gpt-5",            "tokens": 1204, "cost": 0.009632, "latencyMs": 3201, "success": true },
        { "model": "google:gemini-2-pro",     "tokens": 987,  "cost": 0.004935, "latencyMs": 2876, "success": true },
        { "model": "meta:llama-4-maverick",   "tokens": 1108, "cost": 0.002216, "latencyMs": 3540, "success": true }
      ],
      "totalTokens": 4821,
      "totalCost": 0.034210
    }
  }
}
panel: true — raw panel response
model is literally "panel". _meta.panel lists each model that contributed.
json
{
  "id": "panel-1740657123456",
  "object": "chat.completion",
  "model": "panel",
  "choices": [{
    "message": {
      "role": "assistant",
      "content": "**openai:gpt-5**\n\nThe CAP theorem states...\n\n---\n\n**google:gemini-2-pro**\n\nCAP stands for...\n\n---\n\n**meta:llama-4-maverick**\n\nIn distributed systems..."
    }
  }],
  "usage": { "total_tokens": 3299 },
  "_meta": {
    "modelUsed": "panel",
    "routerMode": "panel",
    "cost": 0.016783,
    "tokensUsed": 3299,
    "latency": 4120,
    "panel": {
      "models": [
        { "model": "openai:gpt-5",          "tokens": 1204, "cost": 0.009632, "latencyMs": 3201, "success": true },
        { "model": "google:gemini-2-pro",   "tokens": 987,  "cost": 0.004935, "latencyMs": 2876, "success": true },
        { "model": "meta:llama-4-maverick", "tokens": 1108, "cost": 0.002216, "latencyMs": 3540, "success": true }
      ],
      "totalTokens": 3299,
      "totalCost": 0.016783
    }
  }
}

Requirements

  • At least 3 fallback models must be configured on the router, ideally from different model families
  • Streaming is automatically disabled — debate mode requires all panelist responses before the arbitrator can run
  • Sufficient credits — billing reserves ~4× the normal cost estimate for debate requests

Debate Node in Agent Designer

You can also use debate mode as a node in the agent workflow designer. The Debate node sends the current workflow input to 3 distinct-family models and returns the arbitrated result.

Node config options:

  • System Prompt — optional prompt sent to all panelist models
  • Arbitrator Prompt — custom instructions for the arbitrator
  • Write Output to State Key — optionally store the result in workflow state

Related

Router Configuration Guide

Learn how to set up fallback models and optimization weights.

Cost Optimization Guide

Strategies for managing costs across your AI workload.