Council of Agents

cd /path/to/councils && claude
> Run the ideas council on: I want to build a peer-to-peer local sync engine for notes — no server, CRDTs for conflict resolution. Should I pursue this?
> Run the ideas council on: I have a conjecture that prime gaps follow a log-normal distribution. Is this worth pursuing?
> Run the ideas council on: A policy idea — give every public school student a $500 annual "learning budget" to spend on any educational resource, no restrictions.

cd /path/to/councils && claude
> Run the financial investments council on: NVIDIA (NVDA) at current price — is the AI infrastructure thesis still intact or is this peak cycle?
> Run the financial investments council on: I'm considering a 10% allocation to long-dated US Treasuries as a tail hedge. Evaluate.
> Run the financial investments council on: Uranium miners ETF (URA) — nuclear energy renaissance thesis.

cd /path/to/councils && claude
> Run the code review council on: def divide(a, b): return a/b
> Run the code review council on: [paste a function or file]
> Council: council_code_review.md — review the authentication module in auth/session.py

Stage 1 — Independent
  Each persona answers the question alone, with no visibility into other responses.
  This prevents anchoring — each brings their full expertise without being pulled
  toward a consensus before one has formed.

Stage 2 — Anonymized Peer Review
  Responses are shuffled and labeled A/B/C/D (identity hidden).
  Each persona reviews ALL labeled responses and must:
    1. Rank them
    2. Find at least one flaw in the top-ranked response
    3. Suggest one change to the response closest to their own perspective
  Forced dissent combats sycophancy. Anonymization prevents favoritism.

Stage 3 — Chairman Synthesis
  The orchestrator (not a persona) synthesizes across all responses and reviews.
  Output is structured as:
    ✓ HIGH CONFIDENCE  — where agents agreed
    ⚠ DISPUTED        — genuine disagreements, both sides shown
    ✦ UNIQUE INSIGHTS  — strong points from only one agent
    ? UNRESOLVED       — things the council cannot resolve

councils/
├── CLAUDE.md                                    ← Approach 3: orchestrator brain for native Claude Code
├── council_ideas.md                             ← Ideas council config
├── council_financial_investments.md             ← Financial investments council config
├── council_code_review.md                       ← Code review council config
├── run_council.py                               ← Approach 1: Python controls the loop
├── run_council_agents.py                        ← Approach 2: Python launches Claude as orchestrator
│
└── personas/
    │
    │   # Ideas council
    ├── reductionist/
    │   ├── SKILL.md
    │   └── references/musk_algorithm.md
    ├── adjacent-possible/
    │   ├── SKILL.md
    │   └── references/johnson_concepts.md
    ├── devil-advocate/
    │   └── SKILL.md
    ├── remarkability-analyst/
    │   ├── SKILL.md
    │   └── references/remarkability_concepts.md
    ├── systems-thinker/
    │   └── SKILL.md
    │
    │   # Financial investments council
    ├── value-investor/
    │   └── SKILL.md
    ├── quant-analyst/
    │   └── SKILL.md
    ├── macro-strategist/
    │   └── SKILL.md
    ├── risk-skeptic/
    │   └── SKILL.md
    └── tail-risk-antifragility/
        ├── SKILL.md
        └── references/taleb_concepts.md
    │
    │   # Code review council
    ├── security-auditor/
    │   └── SKILL.md
    ├── performance-engineer/
    │   └── SKILL.md
    ├── maintainability-critic/
    │   └── SKILL.md
    └── domain-expert/
        └── SKILL.md

Python
  ├── calls Anthropic API → persona A response
  ├── calls Anthropic API → persona B response     Stage 1
  ├── calls Anthropic API → persona C response
  │
  ├── shuffles labels [A→C, B→A, C→B]              (mechanical, guaranteed)
  ├── calls Anthropic API → persona A review        Stage 2
  ├── calls Anthropic API → persona B review
  ├── calls Anthropic API → persona C review
  │
  └── calls Anthropic API → chairman synthesis      Stage 3

Python
  └── launches Claude orchestrator (Agent SDK)
        └── Claude reads orchestrator system prompt
              ├── Agent tool → persona A subagent (isolated Claude process)
              ├── Agent tool → persona B subagent (isolated Claude process)  Stage 1
              ├── Agent tool → persona C subagent (isolated Claude process)
              │
              ├── Claude shuffles labels (intentional, not mechanical)
              ├── Agent tool → each subagent for review                      Stage 2
              │
              └── Claude synthesizes directly                                Stage 3

User runs: claude (in councils/ directory)
  └── Claude reads CLAUDE.md → becomes council orchestrator
        └── User: "Run the ideas council on: ..."
              ├── Claude reads council_ideas.md → gets persona list
              ├── Claude reads each SKILL.md → gets prompt + tools
              │
              ├── Agent tool → reductionist subagent
              ├── Agent tool → adjacent-possible subagent              Stage 1
              ├── Agent tool → devil-advocate subagent
              │
              ├── Claude assembles anonymized review prompts
              ├── Agent tool → each persona for review                 Stage 2
              │
              └── Claude synthesizes directly                          Stage 3