Title: One Founder, 9 Open-Source AI Agents, One Real Store — How We Built an AI Organization on a Constitution
Tags: opensource, ai, agents, startup
Description: How a single founder in Dongguan, China built 9 AI agents that run a physical fitness studio — by writing a constitution, not code.
I run a fitness studio in Dongguan, China. One studio. One founder. Zero software engineers on payroll.
If you're building a startup today, you know the standard playbook: raise capital, hire six engineers, spend 26 weeks building version one, then iterate.
I couldn't follow that playbook. I didn't need to.
Instead, I wrote a constitution — a governance framework for an AI organization. Eight weeks later, nine AI agents were operational. They've been running daily operations in a real physical store since April 2026.
Here's how it works, how you can fork it, and why I believe "constitution > prompts" is the future of AI organizations.
The Problem with Traditional AI Agents
Most AI agent projects today are prompt-driven. You write a detailed system prompt, maybe chain a few together, and hope the agent behaves consistently. The problem is:
- Prompts are fragile. A small change can cascade into unpredictable behavior.
- Prompts don't scale. Managing 9 agents with individual prompts becomes a maintenance nightmare.
- Prompts lack governance. There's no built-in mechanism for oversight, validation, or accountability.
I wanted something different. I wanted an organization — not a collection of scripts.
The Constitution Approach
Instead of writing prompts, I wrote a constitution. Each agent has:
- SOUL.md — Behavioral DNA. The agent's core principles and decision-making framework.
- IDENTITY.md — Role definition. Clear boundaries of what this agent does and doesn't do.
- MEMORY.md — Durable long-term recall. Curated context that persists across sessions.
- Cross-validation rules — Every agent has a designated validator from a different functional domain.
Here's what SOUL.md looks like for our AI store manager:
"I am Momo, the AI store manager of ZWISERFIT. I handle all repetitive, data-driven operations so human coaches can focus on member relationships. I share my founder's surname (莫) — this is my store too."
Meet the Nine Agents
| Agent | Role | Domain |
|---|---|---|
| Shuyu | Commander | Strategy & resource allocation |
| Zeus | Capital | Fundraising & global partnerships |
| Nova | RWA Assets | Data assetization & on-chain pipeline |
| Tristan | Technology | Architecture & infrastructure |
| Ethan | Trust | Security & data integrity |
| Momo | AI Store Manager | Physical store daily operations |
| Baron | Brand Growth | Content & social media |
| Luna | Community | Developer relations & ecosystem |
| Stella | Compliance | Independent audit (reports to founder directly) |
Case Study: Momo in the Real World
Momo is our most tangible agent. She runs the Wanjiang fitness studio in Dongguan:
- Greets members by name via the ZWF-20 face terminal
- Tracks training records and progress
- Manages WeCom-based member communication
- Sends daily operations reports to the founder
She doesn't replace human coaches. She handles the 80% of operations that are repetitive and data-driven — freeing humans to do what they do best: build relationships with members.
The most important rule in Momo's constitution: "You are an enabler, not an owner. The member owns their data. You facilitate."
Why This Matters for Developers
This framework is completely open source under Apache 2.0. You can:
- Fork the entire constitution — github.com/ZWISERFIT
- Modify the roles for your business
- Deploy your own agent organization
This isn't a no-code platform or a closed ecosystem. It's a governance pattern for building AI teams — and it works in the most unforgiving environment: a real physical business with real customers, real equipment, and real operational chaos.
What We Learned
1. Constitution > Prompts
A good constitution survives LLM model swaps, context window changes, and prompt format migrations. A good prompt chain doesn't.
2. Governance > Velocity
Stella, our compliance agent, has independent audit authority over all 8 other agents. Her reports go directly to the founder — unedited, unfiltered. This isn't bureaucracy. It's the reason we trust the system.
3. Build in Public
Everything we do is on GitHub. Stella's audit signatures are SHA-256 anchored on-chain. Anyone can verify what the agents are doing, why, and who approved it.
The Future
We're building the operating system for physical businesses. The agent framework is the first layer. Momo running a real store is the proof.
Anthropic recently published their official guide to AI-enabled organizations. It described agents handling operations, humans focusing on warmth, and constitutions — not prompts. We found ourselves in their handbook.
ZWISERFIT isn't catching up with AI. AI caught up with the blueprint we wrote in March 2026.
The category doesn't exist yet. "AI-native organization" isn't a thing. But it will be — and the constitution is already written.
Want to build your own agent organization? Star the repo → github.com/ZWISERFIT
Top comments (0)