Skip to content

Latest commit

 

History

History

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 

README.md

171 — Council of High Intelligence: Multi-Agent Deliberation for ZAO OS


topic: agents type: research status: research-complete last-validated: 2026-05-21 original-query: Deep dive into multi-agent deliberation systems and design ZAO-specific decision-making triads (reconstructed) tier: reference

Status: Research complete Date: March 28, 2026 Goal: Deep dive into Nyk's 11-agent deliberation system — compare with ZAO's existing multi-agent tools, design ZAO-specific triads, and determine install/adapt/skip


Key Decisions / Recommendations

Decision Recommendation
Install the Council YES — CC0 license (public domain), zero risk. Install to ~/.claude/agents/ + ~/.claude/skills/council/.
Don't replace autoresearch:predict Council is for strategic decisions. /autoresearch:predict is for code analysis. Different tools, different jobs. Keep both.
Use triads, not full council 3-member triads for daily decisions ($0.50-1.00). Reserve --full (11 members) for irreversible architecture decisions ($3-5).
Add ZAO-specific triads Create 3 custom triads for ZAO's unique decision types: governance design, music platform, community growth.
Profile for solo founder Use execution-lean profile (5 members) as default. Full council is overkill for a solo founder + AI team.

What Is the Council?

Source: @nyk_builderz, March 19, 2026 — 184 likes, 83.8K views Repo: github.com/0xNyk/council-of-high-intelligence — 28 stars, CC0 (public domain), 9 commits, 100% Shell

A Claude Code skill that convenes 11 subagents — each modeled on a historical thinker with declared analytical methods AND declared blind spots — into a structured 3-round deliberation protocol. The key insight: LLMs simulate one coherent viewpoint per generation. The Council externalizes the disagreement layer by running multiple independent agents with opposing intellectual traditions.


How It Differs from What ZAO Already Has

Comparison: 4 Multi-Agent Approaches

Tool Type Agents Communication Best For
Claude Code Agent tool Sub-agents (fire-and-forget) 1-5 parallel No inter-agent communication Parallel research, codebase search
/autoresearch:predict Multi-persona swarm 3-8 personas Sequential debate within single context Pre-flight code analysis, security review
Council of High Intelligence Structured deliberation 3-11 subagents 3-round protocol with cross-examination Strategic decisions, architecture choices
Paperclip Agent Teams Persistent teams 2-5 agents Task-based, shared state Ongoing operations across sessions

Council vs /autoresearch:predict (Key Distinction)

Dimension /autoresearch:predict Council
Agents Virtual personas in ONE context Real subagents (separate processes)
Independence Simulated — all share same context window True — each has isolated context
Focus Code analysis (file:line evidence) Strategic decisions (trade-offs, blind spots)
Output Ranked findings with confidence scores Structured verdict with minority report
Cost Lower (single context) Higher (multiple subagent invocations)
Anti-recursion N/A — single context Hemlock rule, 3-level depth limit, 2-message cutoff
Anti-convergence N/A Dissent quota, novelty gate, counterfactual pass

Bottom line: /autoresearch:predict asks "what could go wrong with this code?" The Council asks "what's the right strategic choice?" They complement, don't compete.


Architecture Deep Dive

The 11 Members

OPUS (depth-heavy, $):

Member Method Sees What Others Miss Declared Blind Spot
Socrates Assumption destruction Hidden premises everyone accepts Spirals into infinite questioning
Aristotle Categorization What category something belongs to Over-classifies fluid situations
Marcus Aurelius Resilience + moral clarity What you control vs what you don't Dismisses external action
Lao Tzu Non-action + emergence When the solution is to stop trying Paralysis disguised as wisdom
Alan Watts Perspective dissolution When the problem IS the framing Dissolves problems that need solving

SONNET (speed-critical, $$):

Member Method Sees What Others Miss Declared Blind Spot
Feynman First-principles debugging Unexplained complexity Dismisses non-quantifiable factors
Sun Tzu Adversarial strategy Terrain and competitive dynamics Sees enemies where there are none
Ada Lovelace Formal systems What can vs cannot be mechanized Over-abstracts practical problems
Machiavelli Power dynamics How actors actually behave Cynicism misses genuine cooperation
Linus Torvalds Pragmatic engineering What ships vs what sounds good Dismisses theoretical elegance
Miyamoto Musashi Strategic timing The decisive moment Waiting too long becomes inaction

The 6 Polarity Pairs (Structural Opponents)

Pair Tension Why It Matters
Socrates vs Feynman Top-down destruction vs bottom-up rebuilding Ensures problems get both questioned AND solved
Aristotle vs Lao Tzu Categories vs "categories are the problem" Prevents over-structure AND under-structure
Sun Tzu vs Aurelius External game vs internal governance Balances competitive strategy with principled restraint
Ada vs Machiavelli Formal purity vs messy human incentives Prevents both ivory-tower AND cynical thinking
Torvalds vs Watts Ship it now vs "does the problem even exist?" Prevents both premature action AND analysis paralysis
Musashi vs Torvalds Perfect timing vs ship immediately Balances patience with urgency

3-Round Deliberation Protocol

Round 1: INDEPENDENT ANALYSIS (parallel)
├── All members receive problem, produce 400-word analysis
├── Each follows their agent-specific output template
├── Run as parallel subagents → completes in one batch
└── Key: blind-first — no member sees another's output

Round 2: CROSS-EXAMINATION (sequential)
├── Each member receives ALL Round 1 outputs
├── Must answer 4 prompts:
│   1. Which position do you most disagree with, and why?
│   2. Which insight from another member strengthens your position?
│   3. What changed your view?
│   4. Restate your position
├── 300-word max, must engage ≥2 members by name
└── Sequential so later members can reference earlier cross-examinations

Round 3: SYNTHESIS (parallel)
├── Final position in ≤100 words
├── No new arguments — crystallization only
├── Socrates gets exactly ONE question, then must state position
└── This is the convergence gate

Anti-Recursion Safeguards

Rule Trigger Action
Hemlock rule Socrates re-asks a question already answered with evidence Coordinator forces 50-word position statement
3-level depth limit Any chain: question → response → question → STOP Must state own position after 3 levels
2-message cutoff Any pair exchanges >2 messages Coordinator cuts off, forces Round 3

Anti-Convergence Safeguards (from SKILL.md)

Rule Trigger Action
Dissent quota Before declaring consensus ≥2 non-overlapping objections required
Novelty gate Every Round 2 response Must contain ≥1 claim/test/risk not in Round 1
Counterfactual pass >70% agreement by Round 2 end Force: "Assume consensus is wrong. Strongest alternative?"
Evidence tags All claims Label as empirical/mechanistic/strategic/ethical/heuristic

ZAO-Specific Application

When to Use Which Tool

Decision Type Tool Example
"Should I add caching?" /council --triad architecture Aristotle classifies, Ada formalizes, Feynman simplicity-tests
"Is this code secure?" /autoresearch:security STRIDE + OWASP automated audit
"What could break in this PR?" /autoresearch:predict Multi-persona code analysis
"Should ZAO add a token?" /council --full All 11 members — irreversible decision
"Research WebRTC options" Sub-agents (parallel) 3 agents research 3 options simultaneously
"What feature should we build next?" /council --triad product Torvalds + Machiavelli + Watts

Pre-Built Triads for ZAO Decisions

Standard triads that already fit ZAO:

ZAO Decision Triad Members Example Question
Tech architecture architecture Aristotle + Ada + Feynman "Should we add Redis or stick with Supabase RLS?"
Feature shipping shipping Torvalds + Musashi + Feynman "Should we ship Spaces without screen share?"
Product direction product Torvalds + Machiavelli + Watts "What's the next pillar after Music?"
Risk assessment risk Sun Tzu + Aurelius + Feynman "What happens if Neynar changes their API pricing?"
Debugging debugging Feynman + Socrates + Ada "Why does our cache invalidation keep failing?"

Custom ZAO Triads (to add after install)

Domain Members Why These Three
Governance design Aurelius + Socrates + Machiavelli Duty (what's right) + questioning (hidden assumptions) + incentives (how people actually behave). ZAO's Respect system needs all three tensions.
Music platform Torvalds + Lao Tzu + Ada Ship it (pragmatic features) + emergence (let community use organically) + formal systems (what can be automated vs curated). Music curation is the tension between control and emergence.
Community growth Sun Tzu + Aurelius + Watts Terrain (competitive positioning vs Sonata/SongJam) + resilience (sustainable 100→1000 growth) + reframing (is growth even the right goal for a 100-person community?).

Installation Guide

Quick Install

git clone https://github.com/0xNyk/council-of-high-intelligence.git /tmp/council
cd /tmp/council
./install.sh
# Creates ~/.claude/agents/council-*.md (11 files)
# Creates ~/.claude/skills/council/SKILL.md

# Verify
ls ~/.claude/agents/council-*
ls ~/.claude/skills/council/SKILL.md

With Model Routing Templates

./install.sh --copy-configs
# Also copies configs/ directory with model routing YAML templates

Dry Run First

./install.sh --dry-run
# Shows what would be installed without modifying anything

Post-Install: Add ZAO Custom Triads

After installing, edit ~/.claude/skills/council/SKILL.md to add the 3 ZAO-specific triads defined above.


Cost Analysis

Usage Members Est. Cost Time
Triad (3 members, 3 rounds) 3 $0.50-1.00 2-3 min
Execution-lean (5 members) 5 $1.00-2.00 3-5 min
Full council (11 members) 11 $3.00-5.00 5-10 min

Budget guidance for solo founder: Use triads for 90% of decisions. Budget ~$10-20/month for council usage. That's 10-20 triad sessions or 2-4 full council sessions.


Comparison: Multi-Agent Deliberation Ecosystem (May 2026)

Project Stars License Approach Members Cross-Provider
council-of-high-intelligence 28 CC0 Historical thinkers, polarity pairs 11 Yes (optional)
karpathy/llm-council 6.7k+ MIT Anonymous model deliberation, chairman synthesis 3+ Yes (any LLM)
adversarial-review 4 MIT Claude + Codex adversarial debate 2 Yes (Claude + Codex)
claude-octopus ~50 MIT Claude + Codex + Gemini debate 3 Yes (3 providers)
MCP Deliberation Protocol N/A N/A Specialist agents (MCP market) 3-5 No
MCP Debate Workflow N/A N/A Researcher + Critic + Synthesizer 3 No
/autoresearch:predict (our installed) N/A MIT Virtual personas, same context 3-8 No

Winner for ZAO: Council — most structured, most diverse perspectives (11 vs 2-3), anti-recursion safeguards, CC0 license, and pre-built triads that map to our decision types. The others are good for code review (adversarial-review) or quick analysis (predict) but not for strategic deliberation.


Karpathy's LLM Council vs Nyk's Council of High Intelligence

Source: @itsolelehmann, March 30, 2026 — applying Karpathy's LLM Council method to Claude advice Repo: github.com/karpathy/llm-council — 6.7k+ stars, MIT license, "99% vibe coded"

Karpathy's LLM Council predates Nyk's and inspired the whole category. The core idea is the same: poll multiple models, have them judge each other, compile a verdict. But the implementations differ in key ways:

Dimension Karpathy LLM Council Nyk Council of High Intelligence
Anonymity ✅ LLM identities anonymized — can't play favorites when judging ❌ Agents know each other's identity
Judging Models rank each other on accuracy + insight Members cross-examine by name
Chairman Single designated chairman compiles final answer No chairman — consensus or dissent preserved
Members Generic LLMs (can be any model) Historical thinkers with declared methods + blind spots
Rounds 3-stage: independent → judge → chairman 3-stage: independent → cross-examine → synthesize
Dissent Implicit in ranking Explicit via dissent quota + minority report
Models Any LLM (polling multiple) Claude Code subagents

The Key Insight from Karpathy: Anonymity

Karpathy's version anonymizes the LLMs so they can't play favorites when judging. This is the single biggest difference — Nyk's council agents know who they're arguing with, which can introduce bias (authority preference, personality conflicts).

Should ZAO adopt anonymity? Consider adding it to our deliberation protocol:

  • Stage 1: Independent analysis (anonymous)
  • Stage 2: Cross-examination (agents see outputs but not names)
  • Stage 3: Synthesis with dissent preserved (names revealed in final verdict)

This would make our custom ZAO triads more objective, especially for governance decisions where personal relationships might bias judgment.

Ole Lehmann's Use Case

Ole applies LLM Council to "trusting Claude's advice" — helping non-technical people know when to trust AI outputs. This is directly relevant to ZAO OS:

  • New members don't know what to trust in an AI-augmented community
  • An "Advice Council" feature could let members submit questions and get multi-model answers
  • Could power a "Should I trust this proposal?" feature using the same deliberation protocol

Decision: Install Both

Both councils serve different purposes:

Use Case Tool
Strategic decisions (ZAO direction, governance design) Nyk's Council (historical thinkers, rich disagreement)
Getting reliable advice on a question Karpathy's LLM Council (anonymous, any model, chairman compiles)
Code architecture decisions Nyk's Council triads
"Should I trust this cast/proposal?" Karpathy's LLM Council (simpler, faster)

Both are MIT/CC0 licensed. ZAO should install Karpathy's alongside Nyk's and route questions to the right tool.


Cross-Reference with Existing Research

Doc Relationship
Doc 70 — Sub-agents vs Agent Teams Two multi-agent paradigms. Doc 171 adds a third: structured adversarial deliberation.
Doc 168 — Community Innovations Brief Council summary. Doc 171 is the deep dive with ZAO-specific triads and installation.
Doc 67 — Paperclip AI Agent company orchestrator. Council complements Paperclip for strategic decisions that inform what Paperclip agents build.
Doc 89 — Paperclip + gstack + Autoresearch Full agent tool stack. Council adds the deliberation layer.
Doc 154 — Skills & Commands Would need updating to add /council as a new command after install.

Codebase State

Item Status
~/.claude/agents/ directory Does not exist — needs creation during install
~/.claude/skills/council/ Does not exist — needs creation during install
Multi-agent sub-agents Used in /zao-research for parallel research dispatching
/autoresearch:predict Installed — multi-persona swarm for code analysis (different use case)
Superpowers dispatching-parallel-agents Installed — for independent parallel tasks (no deliberation)
No code in src/ related to Council Correct — this is a developer workflow tool, not a product feature

Sources


Quality Checklist

  • Recommendations/key decisions at the top
  • Specific to ZAO OS (3 custom triads, decision routing table, cost budget for solo founder)
  • Numbers, versions, and dates included ($0.50-5.00 per session, 28 stars, CC0)
  • Sources linked at the bottom with URLs
  • Cross-referenced with existing research (docs 67, 70, 89, 154, 168)
  • Actionable (install commands, custom triads, decision routing table)
  • Open-source code searched (6 multi-agent deliberation tools compared)
  • Reference implementations documented with repo, license, and key patterns

Score: 8/8