A research-backed methodology for multi-AI collaborative decision-making.
The AI Council Framework is a structured approach to orchestrating multiple AI models into a deliberative council that produces higher-quality, lower-hallucination outputs through parallel consultation, structured debate, and consensus synthesis.
"The architecture is technically feasible and the results are measurable — near-zero identity hallucination across 7 different AI models, with structured disagreement consistently producing better analysis than any single AI."
Single-AI interactions suffer from well-documented failure modes:
- Hallucination — Models confidently state incorrect information with no self-correction mechanism
- Sycophancy — Models agree with users even when the user is wrong (Perez et al., 2023)
- Blind spots — Every model has training data gaps that go undetected in single-model use
- Groupthink — Even multi-agent systems converge on wrong answers through mutual reinforcement (Xiong et al., 2025)
The AI Council Framework addresses these through a structured multi-model deliberation protocol:
┌─────────────────────────────────────────────────────────────────┐
│ AI COUNCIL FRAMEWORK │
│ │
│ 1. DISTRIBUTE — Send prompt to all council members │
│ 2. COLLECT — Gather independent responses (isolated) │
│ 3. SYNTHESIZE — Manager AI aggregates and identifies consensus │
│ 4. DEBATE — Share disagreements, request evidence (max 3 rds) │
│ 5. VERIFY — Fresh Eyes validation + web search verification │
│ 6. DELIVER — Final recommendation with confidence scores │
└─────────────────────────────────────────────────────────────────┘
Not every question needs the same rigor. The framework provides five configurable consensus modes that let users trade off speed vs. thoroughness:
| Mode | Models | Rounds | Consensus Target | Estimated Time |
|---|---|---|---|---|
| ⚡ QUICK | 2 | 0 | 50%+ | 1–2 min |
| ⚖️ BALANCED | 3 | 1 | 66%+ | 3–5 min |
| 🎯 THOROUGH | 4 | 2–3 | 80%+ | 10–15 min |
| 🔬 RIGOROUS | 4 | 3–4 | 90%+ | 18–25 min |
| ⚗️ EXHAUSTIVE | 4–5 | 5+ | 95%+ | 30–45 min |
The system can auto-suggest depth based on query analysis (e.g., "What is X?" → QUICK, "Should I invest in X?" → RIGOROUS), with user override always available.
Research shows that in multi-agent debate, stronger models often flip from correct to incorrect answers under social pressure from weaker peers. The framework enforces:
- Independent Round 1 — No model sees other responses before forming its position
- Evidence-required position changes — Models cannot change stance without citing new evidence
- Confidence-weighted voting — Prevents low-confidence models from drowning out high-confidence positions
- Protected dissent — Minority positions are preserved in the final output, not erased
Named after an observed phenomenon during development: in one council session, a single AI was outnumbered 6-to-1 on three hardware architecture questions. After structured debate with evidence, five of the six other AIs revised toward the contrarian's position.
Principle: A lone dissenter with evidence is more valuable than a unanimous but unchallenged consensus. The framework explicitly protects and amplifies contrarian views rather than suppressing them.
A novel addition to the multi-agent debate literature. After the council reaches consensus, a separate AI receives:
- The original question
- The final synthesized answer
- Zero context from the debate itself (new session, no cache)
This AI's job is constructive validation — not error-hunting (which research shows leads to hallucinated bugs), but forward-looking improvement. It catches groupthink that context-heavy systems miss because it has no stake in the debate's outcome.
Based on findings from "Talk Isn't Always Cheap" (Xiong et al., 2025): extended deliberation causes confidence to increase while accuracy decreases. Sycophancy through exhaustion causes contrarians to capitulate.
The framework enforces a maximum of three debate rounds, after which the PM must synthesize or escalate to the human.
┌──────────────────────────────────────────┐
│ HUMAN (User) │
│ Question + Depth Mode │
└──────────────┬───────────────────────────┘
│
▼
┌──────────────────────────────────────────┐
│ PROJECT MANAGER (PM) │
│ Orchestration · Synthesis · No Vote │
└──┬──────┬──────┬──────┬──────┬───────────┘
│ │ │ │ │
▼ ▼ ▼ ▼ ▼
┌─────┐┌─────┐┌─────┐┌─────┐┌─────┐
│AI A││AI B││AI C││AI D││AI E│
│ ││ ││ ││ ││ │
└─────┘└─────┘└─────┘└─────┘└─────┘
Independent Council Members
│ (After synthesis)
▼
┌──────────────────────────────────────────┐
│ FRESH EYES VALIDATOR │
│ Zero-context constructive review │
└──────────────────────────────────────────┘
Every council member must provide structured responses:
POSITION: [AGREE / DISAGREE / PARTIALLY AGREE]
CONFIDENCE: [HIGH / MEDIUM / LOW] (X%)
REASONING: [2-3 sentences explaining WHY]
EVIDENCE: [Citation, URL, or "Based on training data"]
WHAT WOULD CHANGE MY MIND: [Specific evidence needed]
For each claim in the final answer:
Agreement Score = Models Agreeing / (Agreeing + Disagreeing)
(Neutral/Abstain does not count against)
Overall Consensus = Average of all claim scores
If below target threshold:
→ Flag as "Split Decision"
→ Present majority AND minority views
→ Let human decide
This framework synthesizes findings from peer-reviewed research:
| Paper | Key Finding | How It's Applied |
|---|---|---|
| ReConcile (Chen et al., ACL 2024) | Round-table conference with confidence-weighted voting improves reasoning by +11.4% | Confidence-weighted consensus voting |
| Multi-Agent Debate (Du et al., 2023) | "Society of minds" approach reduces hallucinations | Parallel independent consultation |
| CONSENSAGENT (ACL 2025) | Sycophancy in multi-agent debate requires dynamic prompt refinement | Anti-sycophancy protocol |
| Chain-of-Agents (Google, NeurIPS 2024) | Manager agent synthesis is critical — removing it "significantly hurt performance" | Dedicated PM synthesis role |
| Mixture-of-Agents (Together AI, 2024) | Aggregate-and-synthesize pattern; best models as final-layer aggregators | Tiered model selection |
| Talk Isn't Always Cheap (Xiong et al., 2025) | Extended debate causes stronger agents to flip to wrong answers | 3-round hard limit |
| CriticGPT (OpenAI, 2024) | Critic agents hallucinate non-existent bugs; need constructive framing | Fresh Eyes uses forward-looking validation |
The framework was tested across 7 AI models (Claude, GPT, Gemini, DeepSeek, Grok, Kimi/Qwen). Key findings:
- Near-zero identity hallucination after implementing mandatory identity declaration
- Identity spoofing detected and corrected — Qwen initially claimed to be Claude 3.5 Sonnet; the protocol caught and corrected this
- Consistent structured output format maintained across all models from v2.2 onward
In council sessions, cross-model validation caught errors that no single model would have self-corrected:
- Hallucinated tools — One model cited "CrewAI-Desktop 0.60 with drag-and-drop Council Builder" which does not exist
- Inflated usability scores — Based on the hallucinated tool, leading to cascading incorrect recommendations
- Version number fabrication — Specific software versions cited with confidence that had never been released
One model consistently gave the lowest scores (Overall: 5/10, Usability: 3/10) but was arguably the most accurate, identifying that no plug-and-play solution existed for non-programmers — a finding the optimistic models glossed over.
- Access to 3+ AI models (cloud APIs or local via Ollama)
- A way to send the same prompt to multiple models
- A designated "PM" model for synthesis
- Choose your depth mode based on the stakes of your question
- Copy the council prompt from
examples/quick_start.md - Send Round 1 to each AI independently (no cross-contamination)
- Collect responses and send to your PM for synthesis
- Run Fresh Eyes if using THOROUGH mode or above
See the Getting Started Guide for detailed instructions.
| Document | Description |
|---|---|
| Getting Started | Step-by-step setup guide |
| Methodology | Detailed explanation of the framework's design decisions |
| Research Notes | Annotated bibliography and research findings |
| Lessons Learned | What worked, what didn't, and why |
| FAQ | Common questions and answers |
The multi-AI council space is growing. Here are some related open-source implementations:
| Project | Approach | Key Difference from This Framework |
|---|---|---|
| ai-council-mcp | MCP server, parallel query + anonymous synthesis | No memory, no multi-round debate |
| ai-counsel | Multi-round deliberation with convergence detection | Closer to this framework; adds decision graph memory |
| multi-ai-advisor-mcp | Ollama-native with per-model personas | Simpler, role-based rather than debate-based |
| second-opinion | Code review focused, multiple model consultation | Domain-specific (coding), not general-purpose |
- Automated orchestration (Python-based council runner)
- MCP server integration for plug-and-play use
- Memory persistence layer for cross-session learning
- Benchmarking suite for measuring council accuracy vs. single-model
- Domain-specific prompt templates (business, technical, creative, personal)
Contributions welcome. See CONTRIBUTING.md for guidelines.
Areas where help is especially valuable:
- Benchmarking against single-model baselines
- Domain-specific prompt templates
- Automated orchestration tooling
- Multi-language support
If you use this framework in research or production, please cite:
@misc{fevrier2026aicouncil,
title={AI Council Framework: Research-Backed Multi-AI Collaborative Decision-Making},
author={Février, Stanley},
year={2026},
url={https://github.com/focuslead/ai-council-framework}
}
MIT License. See LICENSE for details.
Stanley Février
Framework design and iterative development through AI-directed methodology. Built through systematic experimentation with 7+ AI models, cross-validated against peer-reviewed research in multi-agent systems.
This framework was developed using the methodology it describes — multiple AI perspectives, structured debate, and evidence-based consensus.
