Build a Multi‑Agent AI Bot System in Feishu Using OpenClaw

Many users of OpenClaw run all tasks—copywriting, code generation, image creation—through one main agent, which quickly leads to memory bloat, context contamination, and high token costs. The article proposes a multi‑agent solution that isolates each functional role into its own virtual employee while keeping a single bot endpoint.

Why Multi‑Agent?

Memory burden : The main agent’s memory files (USER.md, memory/…) grow huge over time.

Context confusion : Past interactions (e.g., code changes) pollute the context when the agent is asked to write a public article.

High cost : Each conversation reads a large amount of irrelevant background data.

The proposed solution is a single gateway + one bot + multiple Feishu groups , giving each group a physically isolated expert team (its own workspace, memory, and model).

Two Design Patterns

Clone ("分身术") : One bot is added to different groups; bindings route each group to a different agent. Simple to configure, ideal for personal efficiency.

Independent team ("独立团") : Separate bots for research, design, coding, etc. Each has its own avatar and name, providing strong role identity and true multi‑entity collaboration.

Core Concepts of an Agent

Workspace : The agent’s personal office containing files such as SOUL.md (personality) and prompts.

AgentDir : Identity directory with authentication info and model configuration.

Sessions : Private chat history that never mixes with other agents.

This file‑based isolation makes the AI’s behavior more stable, controllable, and gives the workflow a professional intuition.

Step‑by‑Step Configuration

Step 1 – Add a New Agent via CLI

openclaw agents add work \
    --model zai/glm-4.7 \
    --workspace ~/.openclaw/workspace-work

The command creates a new agent named work with its own workspace.

Step 2 – Set Identity (name and emoji)

openclaw agents set-identity --agent work \
    --name "All‑Purpose Assistant" \
    --emoji "🤖"

Step 3 – Write "On‑boarding" Files

In the agent’s workspace create SOUL.md, AGENTS.md and USER.md. These files define the agent’s personality, capabilities, and user information, turning the CLI‑created body into a fully fledged virtual employee.

Step 4 – Bind the Agent to a Feishu Group

Obtain the group conversation ID from Feishu and add a binding entry:

{
  "bindings": [
    {
      "agentId": "work",
      "match": {
        "channel": "feishu",
        "peer": {
          "kind": "group",
          "id": "oc_d46347c35dd403daad7e5df05d08a890"
        }
      }
    }
  ]
}

Set requireMention": false if you want the bot to respond without being @‑mentioned.

Step 5 – Enable Agent‑to‑Agent Communication

In openclaw.json enable the agentToAgent tool and whitelist the agents that may talk to each other:

{
  "tools": {
    "agentToAgent": {
      "enabled": true,
      "allow": ["main", "mulerun", "brainstorm", "writer", "coder"]
    }
  }
}

Now the chief agent (named main) can dispatch tasks to specialized agents such as brainstorm, writer, or coder.

Agent Communication Mechanism

Agents do not shout at each other; they use OpenClaw’s built‑in sessions_send tool, which acts like an internal telephone line. The chief agent receives user commands, decides which specialized agent should handle the request, and routes the instruction via sessions_send. This ensures that if a worker agent stalls, the chief can intervene and keep the workflow alive.

Final Thoughts

With the multi‑agent architecture in place, you have an AI “army” ready to execute complex pipelines. The real power lies not in stacking models but in designing an effective organizational structure—linear pipelines for sequential tasks or parallel dependency graphs for more intricate workflows.