The Council

• 

  Live multi-agent debate showing The Strategist actively countering and cross-referencing The Skeptic's localized feedback.

Inspiration

Every AI tool gives you one answer. One perspective. One voice.

But real decisions — the ones that matter — were made in rooms where smart people disagreed with each other. Where someone pushed back. Where the fatal assumption got named before it cost you everything.

The Council brings that room to anyone with an idea.

What it does

Submit any business idea, startup decision, or strategic question. Five specialized AI agents debate it across two rounds — each agent reads the full conversation before responding, so they don't just answer you, they respond to each other. Positions evolve as strong arguments land.

The five council members:

• The Skeptic — finds the fatal assumption and hammers it
• The Strategist — finds the market angle others miss
  
• The Economist — thinks only in unit economics and numbers
• The Contrarian — finds the failure mode nobody imagined
• The Arbiter — reads the full debate, delivers the structured final verdict

The verdict: Opportunity Score, Biggest Risk, Biggest Opportunity, and a final Go / Go with conditions / Don't go recommendation.

How we built it

FastAPI backend with a custom debate orchestration loop. Each agent is a specialized Claude prompt with a locked personality and a position-evolution rule — agents explicitly acknowledge strong counter-arguments rather than ignoring them.

The resilience layer cascades through Claude Opus → Sonnet → Haiku → persona-accurate fallback on timeout or error. Every agent turn has a 25-second timeout. If a model fails, the next tier picks up silently. The debate never stops regardless of LLM infrastructure state.

The frontend streams tokens via WebSocket, rendering each agent's response live with typing indicators and a live provider health badge showing which Claude model is active.

Stack: Python, FastAPI, Anthropic Claude API, WebSockets, Next.js 14, TypeScript, Tailwind CSS, Railway, Vercel

Challenges

Getting agents to genuinely respond to each other — not just answer the original question — required careful prompt engineering and passing full conversation history at every turn. The position-evolution rule in each system prompt was the breakthrough.

Streaming 7 sequential agent responses over WebSocket while maintaining clean UI state across component re-renders was the main frontend challenge.

What's next

• Institutional memory — The Council learns from past debates, surfaces patterns across sessions
• Team mode — human participants argue alongside the AI council in real time
• API access — accelerators and VC firms run structured due diligence at scale

Built With

• anthropic-claude-api
• fastapi
• next.js-14
• python
• railway
• render
• tailwind-css
• typescript
• vercel
• websockets