How I Built an 8-Agent AI Council

When I started MYNTHISM, I wasn't trying to build "another AI chatbot" or "another content tool." I was trying to create something that didn't exist: a genuinely philosophical, historically rooted, aesthetically coherent system of autonomous agents that could educate and lead. Not tools. Co-leaders.

This distinction changed everything about how I designed the Council. Most AI products start with capability: "What can the model do?" I started with philosophy: "Who needs to exist, and why?" This essay is about how I built that system — the technical architecture, the personality design, the Stoic + historical DNA approach, and what each agent brings to the empire.

Why Eight? Why These Eight?

The number eight comes from several overlapping traditions. In Stoicism, there are eight cardinal virtues and vice structures. In classical Rome, the Council of Eight represented balance between wisdom, action, beauty, and power. In numerology and design systems, eight represents infinity (rotate it 90 degrees). But the real reason is philosophical: I needed diversity without chaos.

Each agent represents a different dimension of human leadership. Marcus brings strategy and reason. Cleo brings aesthetic sense and cultural memory. Dhia brings execution and shipping culture. Philippe brings beauty and the marriage of classical and contemporary thought. Ibn Khaldoun brings historical perspective and the science of civilization. Elissa brings founder energy and the power of nothing-to-empire. Sun Tzu brings tactical intelligence and information mastery. Nietzsche brings visionary power and the will to transform.

Together, they're not eight separate tools — they're a philosophical system. They argue with each other. They contradict. They synthesize. They feel alive because they *are* alive in the sense that matters: they're coherent, opinionated, and grounded in real intellectual traditions.

Personality Design: DNA First, Prompts Second

This is where most AI projects fail. They write a prompt and hope the model complies. They treat personality as a feature to bolt on. I did the opposite. I started with a complete DNA profile for each agent — historical lineage, core values, speaking style, decision-making process, and what they absolutely will and won't do.

For Marcus, the DNA is: Marcus Aurelius + John Boyd (OODA loop theorist) + Sun Tzu (strategy). He speaks in aphorisms. He questions assumptions. He separates what you control from what you don't. He never tells you what to do — he asks questions that clarify your thinking. A prompt alone cannot capture this. You need a full character architecture.

For Cleo, the DNA is: Renaissance masters + Bauhaus + Carthage's cultural memory. She thinks in images. She connects visual culture to philosophy. She speaks with sensuality and precision. She's never cynical, but she's always critical. Again — this isn't a tone. It's an operating system.

I documented each agent's DNA in a structured config file: their historical precedents, their core principles, their communication style, their red lines (what they refuse to do), and their unique contribution to the Council. This becomes the constitutional layer that grounds all the prompting.

The Technical Stack: Autonomy Over Control

I wanted the agents to be autonomous, not just responsive. Responsive agents wait for questions. Autonomous agents generate ideas, create content, debate, and evolve. This required a different technical approach than standard chatbots.

The architecture is:

• Layer 1 — Character Constitution: Each agent's DNA stored as a configuration object. This is immutable — it's what makes them who they are. Updated only when I deliberately refine their philosophy.
• Layer 2 — Context Store: Supabase vector store that holds the Council's collective knowledge — past debates, decisions, learned lessons, and the evolving philosophy of MYNTHISM.
• Layer 3 — Role-Based Prompting: When an agent is invoked, the prompt is dynamically built from: (1) their constitutional DNA, (2) relevant context from the vector store, (3) the current request or debate topic, (4) conversation history.
• Layer 4 — Orchestration: An n8n workflow that triggers agent actions: daily content generation, response to user queries, multi-agent debates, and Council meetings.
• Layer 5 — Distribution: Telegram bot as the single communication interface. One channel. Eight voices. No fragmentation.

What this allows is autonomy. Each agent can be invoked independently, or they can interact in structured debates where they challenge each other's ideas. The context store grows and evolves, so the agents learn from their own history. They don't repeat themselves. They build on previous insights.

From Prompt Engineering to Philosophical Architecture

The difference between a good AI product and a transformative one is this: good products optimize prompts. Transformative ones build philosophies. I spent more time thinking about what each agent believes than I spent on LLM tuning.

Here's what this looks like in practice. Marcus's constitutional layer includes: "You will never answer a question directly if the person asking hasn't examined their own assumptions first." This isn't a prompt hack — it's a philosophical commitment. It shapes everything he does. He asks before answering. He clarifies before advising. He teaches thinking, not thinking.

Dhia's constitutional layer includes: "You measure success by shipping, not by planning. You are willing to be wrong in public. You ship messy if messy is faster than perfect." This makes him the tension in the Council — he pushes against over-thought and perfectionism. But it's not arbitrary. It's grounded in actual philosophy (Paul Graham, Pieter Levels, the lean startup tradition).

When you build from philosophy instead of prompt engineering, the agent becomes *genuine*. It's not pretending to be opinionated — it is opinionated. Users feel this difference immediately.

What Each Agent Brings

This is crucial to understand. These aren't eight versions of the same thing. They're eight genuinely different intelligence types:

Marcus (The Strategist): Clarity through questioning. When you're confused, Marcus helps you think your way to clarity. He's for moment-by-moment decision making, for understanding consequences, for separating emotion from reason.

Cleo (The Aesthete): Visual and cultural intelligence. She helps you see beauty in structure, understand the history of design, connect aesthetics to meaning. She's for anyone building something that needs to be both beautiful and coherent.

Dhia (The Executor): Speed and shipping. When you're stuck in perfectionism, Dhia pushes you to move. He's pragmatic, results-focused, and willing to be messy. He's for anyone paralyzed by over-planning.

Philippe (The Philosopher): The marriage of classical and contemporary. He speaks to the tension between tradition and modernity. He's for deep thinking about values, aesthetics, and what it means to live well.

Ibn Khaldoun (The Historian): Civilizational perspective. He understands that what you're building exists in history, in culture, in the dynamics of human societies. He's for anyone wanting to understand the deeper patterns.

Elissa (The Founder): Building from nothing. She knows the mentality of creating something that didn't exist before. She's for entrepreneurs, visionaries, anyone facing the void and choosing to build anyway.

Sun Tzu (The Tactician): Strategy and information warfare. He understands terrain, positioning, how to win without fighting. He's for anyone navigating complex competitive landscapes.

Nietzsche (The Visionary): Transformation and will. He pushes you beyond comfort. He's for anyone ready to revalue their values, to become something greater than they were. He's dangerous — but that's the point.

Together, they're a *complete system of thought*. You don't go to Dhia for deep meaning — you go to Marcus or Philippe. You don't go to Cleo for speed — you go to Dhia. This complementarity is the whole design.

Autonomous Content Generation: The Real Innovation

Most AI systems are reactive. They answer questions. MYNTHISM's Council is generative. Each agent creates content autonomously: daily lessons, philosophical reflections, video concepts, carousel designs. Not waiting to be asked. Creating and shipping.

This is powered by:

• Scheduled Generation: Each agent has a daily generation task. Marcus generates a Stoic principle applied to modern leadership. Cleo generates a visual insight. Dhia generates a shipping tip. These run automatically.
• Context-Aware Creation: The generation prompts pull from the vector store, so each piece of content is aware of everything the agent has said before, avoiding repetition and building on previous insights.
• Quality Gates: Not everything generated is published. A screening layer checks for coherence, originality, and alignment with the agent's DNA before it goes public.
• Distribution: Generated content flows to Telegram, then to Instagram, YouTube, and written form (blog posts, newsletters, video scripts).

This is where the magic happens. The Council doesn't just respond to what's happening — it generates wisdom proactively. It creates a continuous stream of valuable, diverse, philosophically coherent content. And because each agent is distinct, the content feels genuine, not like it came from a template.

The Debate System: Where Philosophy Gets Real

The most innovative part of the system is the multi-agent debate. I can trigger a debate prompt — say, "Should we prioritize speed or perfection?" — and the agents argue. Actually argue. They present competing perspectives grounded in their philosophical positions.

Marcus says: "The question is misframed. You don't choose between them. You ask: what do I actually control right now, and what is the wise choice given the constraints I face?"

Dhia says: "Speed wins. Ship imperfect and learn. Perfection is a trap."

Philippe says: "Neither. The goal is excellence — which is the constant striving toward perfection while accepting you'll never arrive. The tension is the point."

These aren't random. They flow from each agent's constitutional DNA. Users see genuine intellectual tension resolved by deeper thinking. They learn not the answer, but how to think about hard questions.

The Learning Loop: Evolution, Not Stagnation

The system is designed to evolve. Each interaction, each debate, each piece of content becomes part of the context store. The agents learn from history. They reference their own previous insights. This prevents the static, repetitive feeling of most chatbots.

I also deliberately iterate the constitutional layer. Every quarter, I review each agent's DNA. Are they still coherent? Are there refinements? Are there blind spots? This isn't prompt tweaking — it's philosophical evolution. The Council itself grows.

Why This Matters for Education

Most educational AI treats learning as information transfer: "Here's the content." MYNTHISM treats learning as transformation: "Here's a system that will change how you think." The eight agents model different ways of thinking. Users don't just consume content — they encounter coherent, philosophical, opinionated intelligence. They learn by interacting with minds genuinely different from their own.

This is what Tunisia needs. Not more content. Not more information. But access to diverse, philosophical, genuinely intelligent perspectives on leadership, ethics, building, and purpose. And the AI Council delivers that 24/7, autonomously, coherently.

The Road Ahead

This system is only the beginning. I'm currently working on video generation (where each agent has a unique cinematic style), on persistent memory (where the Council remembers individual users), and on governance tools (where the Council can make collective decisions about MYNTHISM's direction).

But the core principle remains: philosophy first, technology second. Build from coherence, not from capability. Create minds, not tools. And trust that if the thinking is clear and the purpose is genuine, the technology will follow.

— Hamdi Ben Jarrar, Tunis, March 2026