Based on an article by "数字生命卡兹克" (Digital Life Kha'Zix) on WeChat, published April 1, 2026. Adapted with additional context.
The Pitch: Run a Company With Nine AI Employees
Here's a sentence nobody expected to write in 2026: a 1,400-year-old Chinese government structure might be the best architecture for coordinating multiple AI agents.
The Three Departments and Six Ministries (三省六部制) was established during the Sui Dynasty around 605 AD. Its original purpose was to split executive power so no single minister could become too powerful. Fast forward to March 2026, and an open-source project called Edict (12.8K stars) has turned this imperial system into a multi-agent orchestration framework.
And someone just used it to build an entire startup in 40 minutes.
How Ancient Bureaucracy Maps to AI Agents
The system works like this. Three "departments" handle the lifecycle of every decision:
The Central Secretariat (中书省) drafts proposals. It's the idea factory — the agent that generates options and strategies.
The Chancellery (门下省) reviews and criticizes. This is the agent whose entire job is to find problems with the proposals. Not to be helpful. Not to agree. Just to poke holes.
The Department of State Affairs (尚书省) takes approved plans and distributes them to six specialized ministries for execution.
The six ministries each own a domain:
- Personnel (吏部) → HR and team planning
- Revenue (户部) → Financial modeling
- Rites (礼部) → Branding and communications
- War (兵部) → Competitive analysis (yes, really)
- Justice (刑部) → Legal compliance
- Works (工部) → Engineering and development
The beauty isn't in any single agent — it's that the structure itself is the prompt. Each agent only thinks from its own perspective. The Chancellery doesn't try to be helpful; it tries to find flaws. Works doesn't care about branding; it cares about whether the tech stack can actually ship.
The 40-Minute Startup
The original article's author decided to test this by giving the system one instruction: build me a product — a minimalist writing app focused on flow state.
Here's what happened:
Minutes 1–5: The State Affairs agent receives the order and immediately delegates. Central Secretariat drafts three competing product concepts. Not one — three. Each with different business models, risk profiles, and target audiences.
Minutes 5–10: The Chancellery tears them apart. One concept gets conditional approval with warnings — "you claim the UI will be so beautiful users will screenshot it for social media, but you've presented zero evidence this assumption is valid." Another concept gets rejected entirely — "you want to add a flow-state analytics dashboard, but flow is inherently anti-analytical. The moment you measure it, it disappears."
The author later admitted this critique changed his mind. He'd been using a writing app with stats tracking and never realized the stats were making him anxious — exactly the opposite of flow.
Minutes 10–30: The six ministries work in parallel. War scans 30+ competitors and concludes the specific combination of features (dynamic backgrounds, white noise, editor, flow timer) has no complete competitor yet — but a V2EX post suggests someone is working on something similar. Time window: six weeks.
Rites names the product "墨沉" (Ink Sinks) — evoking ink dissolving into water. When challenged about the screenshot-sharing assumption, it responds: "Users don't share products. They share the identity of being the person who writes at 3 AM."
Justice flags that individual payment collection codes can't legally receive business income in China, and lists the only fonts cleared for commercial use without licensing fees.
Revenue builds a financial model. Works designs the technical architecture. Personnel drafts a hiring plan.
Minutes 30–40: State Affairs compiles everything into a final PRD. Product name, strategic roadmap, tech stack, financial projections, brand strategy, compliance checklist, staffing plan. Twelve deliverables total.
Forty minutes. For what would take a real team two to three weeks.
Why the Structure Matters More Than the Models
The article makes a point that stuck with us: the author had used multi-agent setups before in Claude Code, but they never felt this impactful. The difference wasn't the AI models — it was the organizational design.
In most multi-agent systems, agents cooperate. They try to help each other. They converge toward agreement. This is polite and efficient and completely misses the point.
The Three Departments system forces structural disagreement. The Chancellery's job is to disagree. It doesn't try to find middle ground. It doesn't offer constructive suggestions. It just says: "This is wrong, and here's why." Then the proposing agent has to deal with that.
This creates something rare in AI systems: genuine multi-perspective analysis. Not "here are three viewpoints" from one model — but three separate agents, each locked into a single perspective, each unable to see the other's blind spots, forced through a process that synthesizes their individual limitations into collective intelligence.
The Bigger Picture
The article ends with a reflection: AI's greatest gift to one-person companies isn't doing more work faster. It's providing the multi-perspective thinking that a solo founder simply can't do alone.
Every solo founder thinks their idea is good. Every solo founder misses the legal edge case, the market timing risk, the brand assumption that sounds clever but doesn't hold up. The Three Departments and Six Ministries system forces each of those blind spots into the open — not through a better model, but through a better process.
A 1,400-year-old solution to the problem of concentrated power turns out to be a surprisingly good solution to the problem of concentrated perspective.
Try It Yourself
- Edict on GitHub — The open-source implementation (12.8K stars)
- All Team & Orchestration projects on Claw4Science →
Source: "数字生命卡兹克" WeChat, April 1, 2026. Adapted by Claw4Science.