Core Concepts
Understand debate modes, agent personas, scoring, convergence, and how AskVerdict's multi-agent system works under the hood.
A deeper look at how AskVerdict's multi-agent debate engine works.
Debate Lifecycle
Every debate follows this lifecycle:
pending → running → completed
↘ failed- Pending — Request received, queued for processing
- Running — Agents are actively debating
- Completed — Verdict delivered successfully
- Failed — An error occurred (provider timeout, invalid input, etc.)
Debate Modes
Fast Mode (1 credit)
Single-round debate. Each agent makes one argument, the judge delivers a verdict. Best for:
- Simple yes/no decisions
- Quick sanity checks
- High-volume batch processing
Balanced Mode (3 credits)
Multi-round debate with cross-examination. Agents argue, rebut, and the judge weighs all evidence. Best for:
- Important business decisions
- Comparing multiple options
- When you need confidence in the reasoning
Thorough Mode (8 credits)
Extended debate with research, multiple rounds, and deep analysis. Includes convergence scoring to measure how much agents agree. Best for:
- Critical strategic decisions
- Complex multi-factor analysis
- Investment or hiring decisions
Agent Personas
Pro Agent
Builds the strongest possible case for the proposition. Presents evidence, analogies, and logical arguments supporting the affirmative position.
Con Agent
Builds the strongest possible case against the proposition. Identifies risks, counterarguments, edge cases, and potential downsides.
Judge Agent
Neutral arbiter that:
- Weighs the strength of each argument
- Identifies logical fallacies
- Considers evidence quality
- Delivers a structured verdict with confidence score
Scoring & Confidence
The confidence score (0–100) reflects how clear-cut the decision is:
| Score | Meaning |
|---|---|
| 80–100 | Strong consensus — one side clearly stronger |
| 60–79 | Moderate lean — arguments favor one side but with caveats |
| 40–59 | Close call — both sides have strong arguments |
| 20–39 | Lean against — but significant counterpoints exist |
| 0–19 | Strong opposition — overwhelming evidence against |
Convergence Scores
In Thorough mode, AskVerdict tracks convergence — how much the Pro and Con agents come to agree over multiple rounds. High convergence suggests the answer is clear; low convergence means genuine uncertainty.
BYOK (Bring Your Own Key)
With a BYOK subscription, you supply your own AI provider keys:
- OpenAI — GPT-4o, GPT-4o-mini
- Anthropic — Claude 3.5 Sonnet, Claude 3 Opus
- Google — Gemini Pro
BYOK keys are passed via the X-Provider-Keys header and are never stored. They exist only in memory for the duration of the request.
curl -X POST https://api.askverdict.ai/v1/verdicts \
-H "Authorization: Bearer vrd_your_api_key" \
-H "X-Provider-Keys: openai:sk-..." \
-H "Content-Type: application/json" \
-d '{"question": "...", "mode": "thorough"}'Credits & Billing
| Plan | Credits | Features |
|---|---|---|
| Free | 10 on signup | All modes |
| Pro | 500/month ($29/mo) | All modes, decision matrix, evidence engine, 25 seats |
| BYOK | Unlimited ($19/mo) | All modes, own API keys |
| BYOK Pro | Unlimited ($39/mo) | Decision matrix, convergence, heatmaps, file upload, 10 seats |
Credit packs are also available: 100 credits ($4.99), 500 credits ($19.99), 2000 credits ($59.99).
Next Steps
- Authentication — API key setup and BYOK configuration
- Create a Verdict — Full endpoint reference
- BYOK Setup Guide — Detailed BYOK configuration
Was this page helpful?