Edict: A Multi-Agent Orchestrator That Borrows a Tang Dynasty Bureaucracy

What Edict Actually Adds to a Multi-Agent Setup

Edict solves a specific problem: multi-agent frameworks often let agents talk to each other and return a result with no oversight, no replayable trail, and no way to stop a bad run. The README calls this the "you get a pile of results you don't know how they were processed" problem. Edict's answer is a fixed pipeline modeled on the Tang dynasty's Three Departments and Six Ministries. The user plays the emperor, a taizi agent sorts incoming messages, then zhongshu plans, menxia reviews and can reject, shangshu dispatches, and six or seven ministries execute in parallel. The target user is someone already running OpenClaw who wants a heavier governance layer on top of agent conversations. The project is not a general-purpose agent library; it is a configuration and dashboard layer for OpenClaw. That distinction matters because the README's comparison table lists CrewAI, MetaGPT, and AutoGen as alternatives, but Edict cannot run without OpenClaw underneath.

The Mandatory Review Stage Is the Core Mechanism

The README's key architectural claim is that menxia, the review department, is not an optional plugin. Every task passes through it before execution. The review checks whether the plan from zhongshu is complete and whether the subtask breakdown makes sense. If the review fails, the task is sent back for rework. This is a hard gate, not a soft warning. The README contrasts this with CrewAI and AutoGen, which it says let agents finish and hand off without a separate quality check. The mechanism is enforced by a permission matrix and by state transition validation in kanban_update.py, which rejects illegal status jumps. That file name appears in the README as the place where valid state paths are enforced. So the review is not just a prompt; it is backed by code that controls the kanban state machine. The trade-off is latency: every task incurs a full plan-review cycle before any execution starts, which is fine for complex tasks but wasteful for trivial ones.

The Dashboard Is a Real Differentiator, Not a Gimmick

The README lists ten dashboard panels, and several of them go beyond what typical agent frameworks show. The kanban view has heartbeat badges for agent health, with green, yellow, and red states. The monitor panel shows task counts per department and a horizontal bar chart. The memorials panel archives completed tasks with a five-stage timeline. There is also a model configuration panel where each agent gets its own LLM, and switching a model restarts the gateway in about five seconds. The skills panel lets you view and add skills per department. The news panel collects tech and finance headlines daily and can push them to Feishu. The court discussion panel runs an LLM-driven multi-role debate among officials. The README's comparison table claims that none of CrewAI, MetaGPT, or AutoGen offer a real-time dashboard, task intervention, or health monitoring. That may be true for those specific frameworks, but the claim is only as strong as the README's own table, which is not independently verified.

Installation Is Script-Driven, but the First Run Has a Catch

The quickest path is Docker: docker run -p 7891:7891 cft0808/sansheng-demo, then open http://localhost:7891. The README warns about an exec format error on x86 machines and suggests adding --platform linux/amd64 or using the provided docker-compose file. The full install requires OpenClaw, Python 3.10+, macOS or Linux, and Node.js 18+ for the frontend build. The install.sh script creates agent workspaces, writes SOUL.md files for each department, registers agents and a permission matrix into openclaw.json, sets up symlinks so all workspaces share the same data directory, enables agent-to-agent communication visibility with sessions.visibility all, and syncs API keys from an existing agent. The catch is that you must first run openclaw agents add taizi and configure an API key before running install.sh again, otherwise the key sync has nothing to copy. The start.sh script runs the data refresh loop and the dashboard server, and there is a systemd service file for production. The README says the dashboard is embedded in server.py, so the Docker image does not need a separate frontend build.

Where Edict Falls Short or Becomes the Wrong Tool

The most obvious limitation is the hard dependency on OpenClaw. If you are not already using OpenClaw, Edict gives you a Docker demo with simulated data, but the real system requires you to manage OpenClaw agents, permissions, and sessions. That is a significant setup burden. The README also documents a platform mismatch issue with the Docker image, which suggests the image is not built for all common architectures. The review gate is a double-edged sword: it slows down every task, so for simple requests like a quick fact lookup or a one-line reply, the full pipeline is overkill. The taizi agent is supposed to filter out casual chat and only create tasks for real commands, but the README does not specify how accurate that filter is. Another limitation is the lack of tagged releases. The repository shows no recent releases, and the last push date is unknown, so you cannot assume stability or a clear upgrade path. The README is heavy on marketing language, like calling the review stage a "killer feature," and the comparison table is self-authored, so treat the feature claims as aspirational until you run the system.

How Edict Compares to CrewAI and AutoGen in Approach

The README explicitly positions Edict against CrewAI, MetaGPT, and AutoGen. The core difference is not the number of agents; it is the presence of a dedicated reviewer with veto power. CrewAI and AutoGen typically let agents collaborate and produce a final output without a separate review step, or they rely on human-in-the-loop only optionally. Edict makes the review mandatory and enforces it through the state machine. AutoGen does support human intervention, but the README claims Edict's review is architectural, not optional. Another difference is observability: Edict's dashboard provides real-time task states, health heartbeats, and a full audit trail of every task's journey. CrewAI and AutoGen do not ship a comparable built-in dashboard, according to the README. The trade-off is that Edict is less flexible: you are locked into a five-stage pipeline, whereas CrewAI and AutoGen let you define arbitrary workflows. If you need a custom collaboration pattern, Edict's fixed bureaucracy will fight you.

Maintenance, Upgrades, and Licensing Considerations

The project is licensed under MIT, which means you can use, modify, and redistribute it freely, but the README does not include a detailed changelog or migration guide. The absence of tagged releases means you cannot pin a version easily, and the install script may change between commits. The README mentions a compatibility agent that supports historical main branches, which suggests the project has evolved and tries to stay backward compatible, but that is a maintenance burden, not a guarantee. The systemd service file and the edict.sh management script indicate a production-oriented design, but they are only as reliable as the underlying OpenClaw setup. The dashboard is embedded in server.py, so upgrading the frontend requires rebuilding the Python server, not just swapping static files. The README recommends running the install script again after adding an API key, which implies that configuration changes are not hot-reloaded for the agents; you must restart the gateway. There is no mention of automated tests or CI in the provided material, so you should run your own smoke tests after any upgrade.

Editorial conclusion

Adopt Edict if you already run OpenClaw and need a structured, auditable pipeline with a mandatory review step and a visual dashboard, especially for tasks where quality control matters more than raw speed. Do not adopt it if you want a standalone framework, a lightweight agent chat, or a system that works without OpenClaw and its agent configuration. Before committing, verify the install script's assumptions about your OpenClaw version, confirm the API key sync works with your provider, and test the docker image on your exact CPU architecture, since the README documents an exec format error on x86 hosts. The project is young, has no tagged releases, and the last push date is unknown, so plan to review the code yourself.

Official sources

1. Official documentation
2. Official README
3. Project repository