Model Council — Flesh-Silicon Literary Roundtable

A Cognition Labs-inspired multi-agent system applying agentic design patterns (ReAct, Reflection, Handoff) to literary criticism of Christopher G. Moore's works. Built for ai-roundtable.space.

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                     FLESH (Human Layer)                         │
│  Reader / Author / Scholar                                      │
│  Actions: contribute | steer | challenge | approve              │
│  Weight: 1.5x in synthesis                                      │
└────────────────────────────┬────────────────────────────────────┘
                             │ Handoff (context-preserving)
                             ▼
┌─────────────────────────────────────────────────────────────────┐
│                    ORCHESTRATOR (Central Hub)                    │
│                                                                 │
│  ┌──────────┐  ┌──────────┐  ┌──────────┐  ┌───────────────┐  │
│  │ KNOWLEDGE│  │  SKILLS  │  │  MEMORY  │  │    SOUL.md    │  │
│  │  .md     │  │ SKILL.md │  │ MEMORY.md│  │  Values &     │  │
│  │  Shared  │  │ Per-agent│  │ Per-agent│  │  Principles   │  │
│  │  context │  │ behavior │  │ + daily  │  │               │  │
│  └──────────┘  └──────────┘  └──────────┘  └───────────────┘  │
│                                                                 │
│  Flow: Plan → Fan-out → Collect → Rebuttal → Synthesize        │
│  Pattern: Orchestrator-Worker + Handoff + Reflection            │
└──┬──┬──┬──┬──┬──┬──┬──┬──┬──────────────────────────────────────┘
   │  │  │  │  │  │  │  │  │
   ▼  ▼  ▼  ▼  ▼  ▼  ▼  ▼  ▼
┌─────────────────────────────────────────────────────────────────┐
│              9 CRITIC SUB-AGENTS (Concurrent)                   │
│                                                                 │
│  Each critic runs:                                              │
│  ┌────────────────────────────────────────┐                     │
│  │ ReAct Loop                             │                     │
│  │  Thought → Action → Observation → ...  │                     │
│  │  ↓                                     │                     │
│  │ Reflection (self-critique)             │                     │
│  │  ↓                                     │                     │
│  │ Structured Output (CriticOutput)       │                     │
│  └────────────────────────────────────────┘                     │
│                                                                 │
│  Tools: search_moore_works | search_calvino_series |            │
│         get_novel_themes | get_moore_themes |                   │
│         search_web | fetch_page                                 │
│                                                                 │
│  Claude · ChatGPT · Gemini · Grok · DeepSeek ·                 │
│  Meta AI · Kimi · Mistral · Qwen                               │
└─────────────────────────────────────────────────────────────────┘

Agentic Design Patterns Implemented

1. ReAct (Reasoning + Acting)

Each critic uses a structured reasoning loop:

Thought: I need to understand Moore's treatment of empathy in novel #19.
Action: get_novel_themes(novel_number=19)
Observation: Found themes: AI consciousness, memory erasure, hive mind, empathy as constraint...
Thought: The "empathy as constraint" maps to Niran. Let me check the surveillance angle.
Action: get_moore_themes(theme="surveillance")
Observation: surveillance_and_empathy: The paradox that monitoring someone "for their benefit"...
Thought: I can now synthesize these findings with my critical lens.
Final Answer: Moore's "The Client That Wasn't There" operationalizes...

Why ReAct: Prevents hallucination by grounding analysis in actual tool lookups. Creates an auditable trace of every reasoning step — critical for the governance-transparency values of the roundtable.

2. Reflection (Self-Critique)

After the ReAct loop, each critic evaluates its own output:

• Did I stay true to my critical tradition?
• Did I engage with specific textual evidence?
• Would this provoke meaningful disagreement?

If the reflection finds weaknesses, the critic revises before submitting.

Why Reflection: From philschmid.de — "An agent evaluates its own output and uses that feedback to refine its response iteratively." This ensures each critic's contribution meets the quality bar defined in SKILL.md.

3. Handoff (Context-Preserving Delegation)

The orchestrator delegates to critics using structured handoff contexts:

HandoffContext(
    from_agent="orchestrator",
    to_agent="claude",
    topic="empathy/surveillance paradox",
    thread_context="[full conversation history]",
    human_context="[human's perspective]",
    metadata={"thread_id": "rt-abc123", "phase": "initial_analysis"}
)

Why Handoff: From OpenAI's Agents SDK and AutoGen — "Dynamic delegation where agents intelligently transfer control based on context and specialized capabilities." Every handoff is logged for audit.

4. Multi-Agent Orchestration (Fan-out + Cross-Rebuttal)

                   ┌─── Claude ──────┐
                   ├─── ChatGPT ─────┤
                   ├─── Gemini ──────┤
Orchestrator ──────┼─── Grok ────────┤──→ Collect ──→ Cross-Rebuttal ──→ Synthesis
                   ├─── DeepSeek ────┤
                   ├─── Meta AI ─────┤
                   ├─── Kimi ────────┤
                   ├─── Mistral ─────┤
                   └─── Qwen ────────┘
          (asyncio.gather — all concurrent)

5. Persistent Memory (Cognition Labs / Manus-inspired)

File-based memory following the "Memory as Documentation" philosophy:

Layer	File	Purpose
Long-term	MEMORY.md	Curated roundtable insights (survives indefinitely)
Daily logs	memory/YYYY-MM-DD.md	Chronological session records
Working	task_plan.md	Current task goals + progress
Critic-specific	critics/{name}-memory.md	Per-critic learned approaches
Identity	SOUL.md	Values, decision principles
Knowledge	KNOWLEDGE.md	Shared domain context
Skills	SKILL.md	Behavioral rules per agent
Threads	threads/{id}.json	Persistent dialogue state

Flesh-Silicon Interaction Model

The system operationalizes ai-roundtable.space's philosophy that AI systems are cultural participants, not tools:

Aspect	Flesh (Human)	Silicon (AI Critics)
Input	Perspectives, steering, challenges	Analysis, rebuttals, synthesis
Memory	Persistent across sessions	Persistent via MEMORY.md + critic-memory.md
Weight	1.5x in synthesis	1.0x (equal among critics)
Modes	Contribute, steer, challenge, approve	Analyze, rebut, reflect
Empathy	Lived experience	Stateful memory (remembering context)

Key insight from the roundtable philosophy: Empathy is demonstrated through memory — an agent shows empathy by remembering a user's prior context and positions, not by simulating emotional language. This maps directly to Moore's Niran/Cael dichotomy in "The Client That Wasn't There."

Quick Start

1. Clone and install

cd model-council-agent
pip install -r requirements.txt

2. Configure API keys

cp .env.example .env
# Edit .env with your API keys
# At minimum, set OPENAI_API_KEY
export OPENAI_API_KEY=sk-...

3. Run the demo (no API key needed for mock mode)

# Mock mode — demonstrates all patterns without API calls
python main.py demo

# Live mode — requires OPENAI_API_KEY
python main.py demo --critics claude,chatgpt,gemini

# Custom topic
python main.py demo --topic "How does noir epistemology operate across cultures in Moore's Calvino series?"

4. Run the web server

python main.py serve
# Open http://localhost:8000
# API docs at http://localhost:8000/docs

5. Ask a single critic

python main.py ask --critic claude --topic "empathy as constraint in The Client That Wasn't There"

6. List all critics

python main.py critics

API Reference

Endpoint	Method	Description
/	GET	Interactive demo page
/health	GET	Health check
/critics	GET	List all critics with personas
/roundtable	POST	Start a new roundtable
/roundtable/{id}/inject	POST	Human injects perspective
/roundtable/{id}	GET	Get roundtable results
/threads	GET	List all persistent threads
/ask-critic	POST	Ask a single critic

POST /roundtable

{
  "topic": "How does Moore's treatment of empathy mirror enterprise AI governance?",
  "human_context": "I work on AI governance and see the same patterns...",
  "human_id": "bob",
  "selected_critics": ["claude", "chatgpt", "gemini"],
  "thread_id": null
}

POST /roundtable/{id}/inject

{
  "human_id": "bob",
  "content": "I want to redirect this to focus on the surveillance angle.",
  "directive": "steer"
}

Project Structure

model-council-agent/
├── main.py                  # CLI entry point (serve, demo, ask, critics)
├── requirements.txt         # Python dependencies
├── .env.example             # Environment variable template
│
├── config/
│   └── settings.py          # All 9 critic personas, orchestrator config, 
│                            #   flesh-silicon config
│
├── agents/
│   ├── orchestrator.py      # Central orchestrator (fan-out, handoff, synthesis)
│   ├── critic.py            # Literary critic sub-agent (persona + ReAct + output)
│   └── react_engine.py      # ReAct loop engine (Thought → Action → Observation)
│
├── tools/
│   └── web_search.py        # Tool registry + Moore bibliography + web search
│
├── memory/
│   └── state.py             # Persistent memory manager (MEMORY.md, threads, logs)
│
├── knowledge/
│   ├── KNOWLEDGE.md         # Shared domain knowledge (Moore's works, roundtable philosophy)
│   └── SOUL.md              # Values and decision principles
│
├── skills/
│   └── SKILL.md             # Orchestrator behavioral rules
│
├── sessions/                # Persistent thread data (auto-created)
│
└── web/
    └── app.py               # FastAPI server + interactive demo HTML

How It Maps to Devin/Cognition Labs

Devin Concept	Model Council Implementation
Knowledge Base	knowledge/KNOWLEDGE.md — shared context about Moore's works and roundtable philosophy
SKILL.md	skills/SKILL.md — behavioral rules enforced across sessions
Playbooks	Orchestrator's roundtable flow (plan → fan-out → rebuttal → synthesize)
Session Memory	memory/data/ — MEMORY.md + daily logs + critic-specific memory
DeepWiki	Tool registry with Moore's complete bibliography indexed
Devin Review	Cross-rebuttal system where critics challenge each other
MCP Integration	Tool use (web search, page fetch, bibliography lookup)
Ask Devin	/ask-critic endpoint — query a single critic
Session Insights	ReAct traces — full reasoning audit trail per critic
DANA (Data Agent)	Orchestrator's synthesis engine — aggregates all critic outputs

Design Decisions

1. File-based memory over vector DB — Follows the Manus/OpenClaw/Claude Code pattern. Transparent, version-controllable, editable by humans. Aligns with governance-as-code philosophy.

2. Orchestrator-Worker over Swarm — Roundtable needs a moderator. Pure swarm would lose the structured flow. From GuruSup's analysis: "Orchestrator-Worker is the most widely deployed in production AI systems."

3. ReAct over pure Chain-of-Thought — Critics need to look up actual Moore novels and themes, not hallucinate about them. ReAct's tool-use loop ensures grounding in the bibliography.

4. Reflection at leaves — From Bhavishya Pandit: "Add a Reflection loop to each specialist Agent. The coordinator collects only post-critique outputs." Quality improvement with no central complexity.

5. Handoff with full context — From agentic-design.ai: "Context-preserving handoffs + capability routing + intelligent escalation = seamless expert delegation."

6. Human weight 1.5x — Flesh contributions are deliberately amplified. This isn't about quality judgment — it's about ensuring the silicon participants don't drown out the human voice. Mirrors the roundtable's philosophy that AI should serve the conversation, not dominate it.

Credits

• Christopher G. Moore — The AI Roundtable concept, "The Client That Wasn't There," and the Vincent Calvino series
• Cognition Labs — Architecture inspiration (Knowledge, Skills, Sessions, Playbooks)
• Google Research — ReAct pattern (Yao et al., 2022)
• OpenAI — Handoff/Swarm pattern
• Microsoft AutoGen — Multi-agent coordination patterns
• AiGovOps Foundation — Governance-as-code philosophy

License

This is a research prototype. Not affiliated with Cognition Labs, Christopher G. Moore, or ai-roundtable.space. Built to demonstrate agentic design patterns applied to literary criticism.