What is this?
Edict is a multi-agent orchestration system for OpenClaw: 12 AI agents (11 business roles plus one compatibility agent) work as an ancient Chinese imperial bureaucracy — planning, review, dispatch and parallel execution — with a real-time dashboard and full audit trails.
Edict is a multi-agent orchestration system for OpenClaw: 12 AI agents (11 business roles plus one compatibility agent) work as an ancient Chinese imperial bureaucracy — planning, review, dispatch and parallel execution — with a real-time dashboard and full audit trails.Is it a framework I code against, or a finished system I just run?A finished system you operate rather than a library you import: the installer creates 12 OpenClaw agent workspaces with persona files, a Python backend (SQLAlchemy + Redis) tracks task state, and a built-in web dashboard is the main interface.
A finished system you operate rather than a library you import: the installer creates 12 OpenClaw agent workspaces with persona files, a Python backend (SQLAlchemy + Redis) tracks task state, and a built-in web dashboard is the main interface.How do I get it running?Quickest look: docker run -p 7891:7891 cft0808/edict, then open port 7891 — the dashboard loads with mock demo data. For real use, install OpenClaw and run ./install.sh; a systemd setup is documented for production.
Quickest look: docker run -p 7891:7891 cft0808/edict, then open port 7891 — the dashboard loads with mock demo data. For real use, install OpenClaw and run ./install.sh; a systemd setup is documented for production.What do I need installed for the full (non-demo) setup?OpenClaw, Python 3.10+ and macOS or Linux. Node.js 18+ is optional — the installer builds the React frontend with it and simply skips the build if it is missing.
OpenClaw, Python 3.10+ and macOS or Linux. Node.js 18+ is optional — the installer builds the React frontend with it and simply skips the build if it is missing.Any install pitfalls I should know about?On x86/amd64 hosts (Ubuntu, WSL2) the image may fail with exec format error — rerun with --platform linux/amd64 or use the bundled docker-compose, which sets the platform. The first install also needs an API key configured (openclaw agents add taizi) before re-running ./install.sh.
On x86/amd64 hosts (Ubuntu, WSL2) the image may fail with exec format error — rerun with --platform linux/amd64 or use the bundled docker-compose, which sets the platform. The first install also needs an API key configured (openclaw agents add taizi) before re-running ./install.sh.Do I need to already be an OpenClaw user?Yes for real deployment — agents are registered into openclaw.json and each gets its own workspace, skills and model. Only the Docker demo runs standalone, serving the dashboard on mock data.
Yes for real deployment — agents are registered into openclaw.json and each gets its own workspace, skills and model. Only the Docker demo runs standalone, serving the dashboard on mock data.Why model AI agents on a 1300-year-old Chinese bureaucracy?The author's argument: most multi-agent frameworks let agents chat freely and hand back results nobody can reproduce, audit or intervene in. The Tang dynasty system builds in separation of powers — a dedicated reviewer department (Menxia) that can formally reject substandard plans and force rework before anything is executed.
The author's argument: most multi-agent frameworks let agents chat freely and hand back results nobody can reproduce, audit or intervene in. The Tang dynasty system builds in separation of powers — a dedicated reviewer department (Menxia) that can formally reject substandard plans and force rework before anything is executed.How does a task actually flow through the system?Taizi (the crown-prince agent) triages incoming messages and only real edicts become tasks. Zhongshu plans and splits them, Menxia reviews and can reject plans into a rework loop, Shangshu dispatches subtasks to the six ministries in parallel, and finished work is archived as a memorial with a five-stage timeline.
Taizi (the crown-prince agent) triages incoming messages and only real edicts become tasks. Zhongshu plans and splits them, Menxia reviews and can reject plans into a rework loop, Shangshu dispatches subtasks to the six ministries in parallel, and finished work is archived as a memorial with a five-stage timeline.What can I control from the dashboard?Ten panels: a Kanban with stop/cancel/resume actions and heartbeat badges, an agent health monitor, archived memorials, 9 preset edict templates with estimated time and cost, a token-consumption leaderboard, per-agent LLM hot-swapping, skills management, live session previews, a news feed with Feishu push, and an LLM-driven court debate panel.
Ten panels: a Kanban with stop/cancel/resume actions and heartbeat badges, an agent health monitor, archived memorials, 9 preset edict templates with estimated time and cost, a token-consumption leaderboard, per-agent LLM hot-swapping, skills management, live session previews, a news feed with Feishu push, and an LLM-driven court debate panel.Can agents just message and override each other freely?No. A strict permission matrix defines exactly who may send messages to whom, and kanbanupdate.py enforces a state machine for task transitions — illegal jumps like Doing→Taizi are rejected and logged, with every state change written to an audit log for full traceability.
No. A strict permission matrix defines exactly who may send messages to whom, and kanbanupdate.py enforces a state machine for task transitions — illegal jumps like Doing→Taizi are rejected and logged, with every state change written to an audit log for full traceability.Which topics is it in?feishu, kanban, event-driven, audit-log, rbac, openclaw, autonomous-agents, workflow-automation, redis, orchestration, agent-orchestration, observability
feishu, kanban, event-driven, audit-log, rbac, openclaw, autonomous-agents, workflow-automation, redis, orchestration, agent-orchestration, observability