cft0808/edict

A multi-agent orchestration system built on OpenClaw that models AI collaboration on the Tang dynasty's "Three Departments and Six Ministries" bureaucracy, adding a mandatory review stage and a real-time dashboard.

What it is

Edict (三省六部) coordinates twelve specialized AI agents through a fixed, government-inspired workflow: a triage agent sorts incoming messages, a planning department decomposes tasks, a review department approves or rejects plans, a dispatch department assigns work, and several ministry agents execute in parallel. It is aimed at people who want multi-agent task execution that is auditable, observable, and interruptible rather than a black box. Compared with frameworks like CrewAI or AutoGen, its distinguishing pieces are an institutional review gate that can reject and force rework, and a real-time kanban dashboard. It requires OpenClaw to run the agents.

Key features

A twelve-agent hierarchy with an explicit permission matrix governing which agent may message which, plus state-transition validation that rejects illegal task status jumps.
A mandatory review department (门下省) that inspects planning output and can reject it back for rework before execution proceeds.
A real-time kanban dashboard with ten panels: task board, dispatch monitor, archived "memorials" with a five-stage timeline, template library, officials/token-usage overview, news feed, per-agent model config, skills management, session monitoring, and a daily ceremony animation.
Per-agent configuration: each agent has its own workspace, skills, and hot-swappable LLM (changing a model restarts the gateway to take effect).
Remote skills management from GitHub/URLs via the dashboard UI, a CLI (skill_manager.py), or a REST API.
An async backend with a Redis Streams EventBus, a transactional Outbox Relay for at-least-once event delivery, a DAG orchestrator, and a parallel dispatch worker with exponential-backoff retry and resource locks; all state changes are written to an audit log.

Tech stack

Python 3.10+ (the README badge shows 3.9+; the install prerequisites state 3.10+); macOS or Linux.
A dashboard server built on the Python standard library (http.server) with zero required dependencies, serving both the API and static files; optional numpy for the LinUCB model-recommendation router and optional Playwright for screenshots.
A React 18 frontend with TypeScript, Vite, and Zustand (13 components), plus a single-file zero-dependency dashboard.html fallback.
An async backend service using SQLAlchemy and Redis (Alembic migrations present under edict/backend).
Runs on OpenClaw (required); ships install.sh/start.sh scripts, a systemd unit, and a Docker demo image with prebuilt frontend and mock data.
MIT licensed.

When to reach for it

Running multi-agent tasks where you need a mandatory quality-review/approval step before work is executed.
Situations that demand full observability and intervention — pausing, cancelling, or resuming tasks and inspecting the complete flow of each one.
Users already running OpenClaw who want a dashboard-driven orchestration layer on top of it.
Trying the concept quickly via the Docker demo image with preloaded sample data.

When not to reach for it

The full system requires OpenClaw and API keys configured per agent; if you are not on OpenClaw it does not apply (only the demo dashboard runs standalone).
The rigid department workflow with a forced review gate adds latency and structure that is unnecessary for simple single-agent or one-shot tasks.
The heavy metaphor (imperial roles, Chinese-first docs) may be a barrier for teams wanting a neutral, plainly named framework.
The full async backend brings Redis and a database into the picture; overkill when you only need lightweight coordination.

Maturity signal

Created in February 2026 and pushed within days of this writing, the project is actively developed and has attracted a large audience. The Phase 1 roadmap is largely complete (twelve-agent architecture, dashboard, memorials, templates, tests), while Phase 2 shows a mix of shipped infrastructure (EventBus, Outbox Relay, state-machine audit, DAG orchestrator) and still-open items such as a human-approval mode. End-to-end tests (17 assertions) and a state-machine consistency test exist. It reads as a young but energetically maintained project whose core is stable and whose institutional features are still deepening.

Alternatives

CrewAI — use it when you want a widely adopted role-based multi-agent framework and do not need a built-in review gate or dashboard.
AutoGen — use it when you want conversational multi-agent orchestration with human-in-the-loop, without the fixed department pipeline.
MetaGPT — use it when you want a company-metaphor multi-agent workflow for software tasks rather than the imperial-bureaucracy model here.
LangGraph — use it when you want to build custom agent graphs and state machines yourself instead of adopting a prescribed hierarchy.

Notes

The dashboard server is deliberately dependency-free: server.py (~2300 lines) runs on http.server and serves both the API and the prebuilt React frontend, and a single-file dashboard.html (~2500 lines) works with no build step.
State transitions are enforced by a _VALID_TRANSITIONS state machine in kanban_update.py; illegal jumps (for example Doing to Taizi) are rejected and logged, and every change is appended to an audit log.
Incoming task titles and notes are automatically cleaned of file paths, metadata, and invalid prefixes before a task is created.

Wiki: cft0808/edict