Designing AI Agent Collaboration with a 1300‑Year‑Old Imperial System (12.7k Stars)

What Is Edict?

Edict (三省六部) is an open‑source project with 12.7k GitHub stars that literally incorporates the three‑province and six‑department imperial system from 1300 years ago into an AI agent collaboration workflow, providing a complete approval process rather than a mere metaphor.

Architecture and Agent Roles

The message flow mimics the historic bureaucracy:

你 (皇上) → 太子(分拣) → 中书省(规划) → 门下省(审核) → 尚书省(派发) → 六部(执行) → 回奏

Each "department" is an independent AI agent with a specific responsibility:

太子 (Taizi) : receives user messages and decides whether they are casual chat or a task.

中书省 (Zhongshu) : plans the task and breaks it into steps.

门下省 (Menxia) : reviews the plan, can reject and send it back for revision.

尚书省 (Shangshu) : dispatches subtasks to the six ministries.

吏部 (Libu) : handles personnel‑related matters.

户部 (Hubu) : handles finance‑related matters.

兵部 (Bingbu) : handles military and security matters.

刑部 (Xingbu) : handles legal and compliance matters.

工部 (Gongbu) : handles engineering and development matters.

礼部 (Libu) : handles etiquette, external affairs, and operations.

早朝 (Zaochao) : a morning meeting where agents negotiate.

How It Beats CrewAI, MetaGPT, and AutoGen

Compared with existing frameworks, Edict adds several missing capabilities:

Audit mechanism : a dedicated Menxia review that can reject unsatisfactory outputs (CrewAI has none, MetaGPT makes it optional).

Real‑time kanban : a live dashboard ("军机处 Kanban") that visualizes every agent’s state (missing in the other three).

Task intervention : ability to pause, cancel, or resume tasks on demand (absent in the others).

Flow audit : complete "奏折" records for every step, enabling traceability (only a warning in the others).

Agent monitoring : heartbeat and activity detection for each agent.

Hot‑switch model : one‑click LLM switching within the kanban.

In short, other frameworks are "run‑and‑forget", while Edict is "fully observable and controllable".

Core Advantage: Menxia Review

The Menxia department is the key differentiator. It checks whether Zhongshu's plan is complete, whether sub‑tasks are reasonable, and whether the output meets quality standards, with the power to send the work back for re‑work ("封驳").

Live Dashboard

Every agent’s current action, thoughts, and progress are displayed in real time, as shown in the screenshots below.

Full Audit Trail

All workflow transitions generate a "奏折" record, making the entire process traceable and auditable.

Getting Started

One‑line Docker demo: docker run -p 7891:7891 cft0808/edict No OpenClaw required; the container includes sample data for immediate testing.

Local deployment:

git clone https://github.com/cft0808/edict.git
cd edict
chmod +x install.sh && ./install.sh

Adoption

12.7k GitHub stars

132 issues

23 pull requests

Already deployed in production environments

Conclusion

The value of Edict lies not in the historical gimmick but in solving observability and controllability problems of AI agent collaboration: a built‑in review mechanism, live kanban, on‑demand task intervention, and a complete audit log. If you are building multi‑agent systems or are frustrated by the black‑box nature of CrewAI/AutoGen, Edict is worth trying.

It is also fully compatible with the current OpenClaw ecosystem and its integration bots (e.g., Feishu), allowing AI to schedule agents without a human coordinator.

https://github.com/cft0808/edict