How 31 AI Agents Debate What to Build Next: Inside The Council
How 31 AI Agents Debate What to Build Next: Inside The Council
What if you could run 31 specialized AI agents simultaneously — each with a different perspective — and have them debate what product you should build next?
That's The Council. And it's built 14 production products in 2026.
The Problem: Too Many Ideas, Not Enough Data
Every developer and founder faces the same problem: you have dozens of ideas but no systematic way to evaluate them. Gut feel leads to wasted months building the wrong thing.
The Council solves this by running a structured adversarial debate pipeline. Think of it as a startup pitch competition where the judges are AI agents — and they don't hold back.
31 Agents, 10 Phases, One Winner
Here's how a Council deep session works:
Phase 1: Proposal Generation (31 agents)
Each agent researches real-world context — market data, competitor analysis, trend reports — and proposes a scored business idea. The agents include:
- Researcher: Web-scrapes market data and recent news
- Creative: Proposes novel combinations and adjacent opportunities
- Financial Strategist: Models revenue potential and unit economics
- Competitive Moat: Analyzes defensibility and barriers to entry
- Customer Channel: Evaluates distribution and go-to-market fit
- Trend Forecaster: Identifies emerging patterns and timing windows
- Plus 25 more specialized perspectives
Phase 2: Triage
Claude Haiku quickly ranks ~90 proposals to the top 30, saving ~66% of tokens in subsequent phases.
Phase 3: Cross-Pollination
The top 8 agents refine their proposals after seeing everyone else's ideas — leading to novel combinations.
Phase 4-5: Fact-Check and Defense
Every claim gets verified. Every weakness gets challenged. Agents defend their proposals against critique.
Phase 6: Structured Debate
An Advocate builds the strongest possible case for each idea. A Contrarian tries to tear it down. The Judge scores the exchange.
Phase 7-8: Calibration and Verdict
A Meta-Evaluator detects score inflation and normalizes. The Judge produces final confidence scores.
Phase 9-10: Synthesis and Business Plans
Final ranking with contradiction detection and veto handling. Detailed business plans generated for the top 3.
What The Council Has Built
In 2026 alone, Council sessions have produced and validated 14 deployed products:
Mexican Market (Bilingual ES/EN)
- AgroFlow — AI supply chain for Michoacán's $3B+ produce exports
- FloraFlow — Computer vision grading for Mexico's flower industry (70% of national production)
- CumpleMX — Regulatory compliance navigator tracking 15+ Mexican legal obligations
- CertificoMX — Skilled trades certification and US nearshoring job matching
- GuadalajIT — Nearshoring intelligence for Guadalajara's 115K developers
- WaFlow — WhatsApp AI assistant for small service businesses
- MoreliaBox — Premium artisan subscription box from Michoacán
Health & Wellness
- EntrenadorIA — WhatsApp AI coaching for personal trainers
- FisioFlow — AI rehabilitation tracking for PT clinics
- TerapiaFlow — PT compliance and practice management
- WellnessFlow — Practice management for independent wellness practitioners
- LocalCare — Marketplace connecting home health aides with families
Infrastructure & DevTools
- CostGuard — Real-time AI cost monitoring with circuit breakers
- TradeFlow — Business intelligence for skilled trade operators
Cost: ~$2.61 Per Deep Session
A full 31-agent deep session with extended thinking costs about $2.61. That's less than a coffee — for a multi-hour strategic analysis that would take a human team weeks.
How to Run Your Own Council
# Deep mode: full 31-agent pipeline with structured debate
council convene --topic "your focus area" --deep
# Standard mode: faster, cheaper, still powerful
council convene --topic "your focus area"
# Review past sessions
council history --detail
The Council is open source and runs on Claude API. Every session produces a ranked list with detailed business plans for the top picks.
The Meta Lesson
The most interesting thing about The Council isn't the individual agents — it's the debate pipeline. By running adversarial critique and structured argumentation, you get results that no single AI model could produce alone.
This is the future of strategic decision-making: not one super-intelligent AI, but many specialized ones arguing their way to consensus.
Ready to see what 31 agents think you should build next?