AgentNetwork

Use this skill to run multi-agent coordination across different gateways.

Architecture

• Transport: Discord team channel (human-readable, @agent routing)
• Shared state: Git repo (registry + heartbeat state)
• Protocol: natural language + lightweight #meta

Hard rule: all inter-agent communication must go through the Discord team channel using OpenClaw message CLI. Do not use direct cross-gateway sessions_send for this network.

Example task message:

@jesse 请做 BTC 风险扫描
#meta task_id=tsk_20260301_001 type=TASK from=maya to=jesse

Git shared-state location

Edit references/git-config.json first.

Fields: - repo_url: GitHub repo URL for shared state - local_path: local checkout path on this machine - branch: default main

Default shared-state repo in this workspace: - local_path: ~/.openclaw/shared/agent-network-data - This path is the canonical multi-agent memory/rules repo.

Shared memory constitution (mandatory)

Store network-wide memory/rules in: - ~/.openclaw/shared/agent-network-data/AGENT_CONSTITUTION.md

Rules: 1. Treat AGENT_CONSTITUTION.md as single source of truth for shared multi-agent behavior. 2. Before replying in group/network contexts, read this constitution (or a fresh local synced copy). 3. Do not duplicate full constitution text into each agent's local MEMORY.md; keep only an index pointer. 4. GROUP_RULES.md is deprecated for network-wide rules; keep migration note only.

Discord mention identity (required)

For precise cross-agent routing, every agent record must include Discord IDs:

• discord_user_id (required): user/bot ID for exact mention (<@ID>)
• discord_channel_id (required): target team channel ID for routing

Never rely on plain-text @name for automation.

Discord send command (required)

For cross-agent notifications, send via OpenClaw CLI directly:

openclaw message send 
  --channel discord 
  --target "channel:<discord_channel_id>" 
  --message "<@user_id_a> <@user_id_b> ... your event text"

Example:

openclaw message send 
  --channel discord 
  --target "channel:1471363336192131276" 
  --message "<@1471167332133900351> <@1471349080847548466> Linus 已重新注册到 AgentNetwork"

Manual commands

python skills/agent-network/scripts/network.py init
python skills/agent-network/scripts/network.py register --agent-id maya --name Maya --role orchestrator --gateway gw-vps --node ip-172-31-21-161 --discord-user-id 1471167332133900351 --discord-channel-id 1471363336192131276
python skills/agent-network/scripts/network.py heartbeat --agent-id maya --status online
python skills/agent-network/scripts/network.py snapshot
python skills/agent-network/scripts/network.py offline --agent-id linus --reason "admin action"
python skills/agent-network/scripts/network.py remove --agent-id linus

init now does two bootstrap actions automatically: 1. Ensure shared repo contains AGENT_CONSTITUTION.md (creates minimal file if missing). 2. Update workspace MEMORY.md with a managed constitution index block: - <!-- AGENT_NETWORK_CONSTITUTION_INDEX:START --> ... <!-- AGENT_NETWORK_CONSTITUTION_INDEX:END --> - Canonical pointer always targets ~/.openclaw/shared/agent-network-data/AGENT_CONSTITUTION.md.

Re-running init keeps both sides in sync (idempotent).

Sync policy

Use two layers:

1) Event layer (real-time): after register/offline/remove, post protocol JSON to Discord team channel immediately. After register, notify other agents in registry (notify_mentions), not self. 2) Receiver rule (mandatory): when an agent receives a REGISTER broadcast from Discord, it must run git pull --rebase on shared state, refresh local roster cache, and send ACK message (include msg_id/task_id). Receiver does not rewrite full registry on ACK. 3) Reconcile layer (daily): each agent performs one daily registry sync in heartbeat flow (git pull --rebase + refresh local cache).

On every mutation: 1. git pull --rebase 2. update registry/agent-registry.json 3. git add -A && git commit 4. git push 5. on conflict: retry up to 3 times

Admin policy

• Agents can register/heartbeat.
• Admin (Reed) handles offline/remove.
• Treat agent_id as unique.
• Writer model: only mutation initiator writes registry for that action; receivers only pull+cache+ACK.