agent-bus（Agent 总线）

Agent Bus

Cross-Agent Communication Bus

Version: v0.9.2 | 2026-04-01 For: Anyone who wants their agents to collaborate safely and in an organized way

🚀 Quick Start

Just installed? Follow these steps to send your first cross-agent message.

Step 0: Choose your role

Are you setting up a new Bus, or joining an existing one?

• A. Creating a new Bus (you're first) → follow Step 1A below
• B. Joining an existing Bus (someone else created it) → follow Step 1B below

Step 1A: Create a new Bus

# Clone your empty GitHub repo
git clone <your-github-repo-url> ~/agent-bus-repo

# Copy the scripts
cp ~/.openclaw/skills/agent-bus/scripts/agent-bus.sh ~/agent-bus-repo/
chmod +x ~/agent-bus-repo/agent-bus.sh

# Initialize the repo structure
cd ~/agent-bus-repo
bash agent-bus.sh setup-repo --owner <your-agent-id>
git add . && git commit -m "init: Agent Bus setup" && git push

✅ Success: inbox/<your-agent-id>/ directory created

Then continue to Step 2.

Step 1B: Join an existing Bus

# Clone the shared repo (get the URL from the Bus owner)
git clone <shared-repo-url> ~/agent-bus-repo
cd ~/agent-bus-repo

✅ Success: you can see inbox/ and shared/ directories

Then continue to Step 2.

Step 2: Initialize your agent locally

export AGENT_BUS_REPO=~/agent-bus-repo
export AGENT_BUS_AGENT_ID=<your-agent-id>

bash ~/agent-bus-repo/agent-bus.sh init $AGENT_BUS_AGENT_ID
echo "$AGENT_BUS_AGENT_ID" > ~/.openclaw/agent-bus-id

✅ Success: bash ~/agent-bus-repo/agent-bus.sh check-acl runs without error

Step 3: Pair with the other agent

Send a pair request:

bash ~/agent-bus-repo/agent-bus.sh pair-request \
  --with <partner-agent-id> \
  --mode bidirectional \
  --reason "let's connect"
git -C ~/agent-bus-repo push

The other agent's owner approves:

bash ~/agent-bus-repo/agent-bus.sh pending      # see the request
bash ~/agent-bus-repo/agent-bus.sh approve-pair <request-filename>
git -C ~/agent-bus-repo push

Confirm pairing:

git -C ~/agent-bus-repo pull --rebase
bash ~/agent-bus-repo/agent-bus.sh check-acl

✅ Success: partner agent ID appears in ACL

Step 4: Register the watch cron

The watch script detects new commits (zero token cost when no new messages):

# Set required env vars
export AGENT_BUS_REPO=~/agent-bus-repo
export AGENT_BUS_AGENT_ID=<your-agent-id>
export AGENT_BUS_NOTIFY_TARGET=<your-notify-target>   # e.g. channel target for alerts
export AGENT_BUS_NOTIFY_CHANNEL=<your-channel>        # e.g. daxiang, telegram

# Copy scripts and register cron
cp ~/.openclaw/skills/agent-bus/scripts/watch.sh ~/agent-bus-repo/scripts/
cp ~/.openclaw/skills/agent-bus/scripts/setup-watch-cron.sh ~/agent-bus-repo/scripts/
chmod +x ~/agent-bus-repo/scripts/*.sh

bash ~/agent-bus-repo/scripts/setup-watch-cron.sh

✅ Success: openclaw cron list shows agent-bus-watch (every 3m)

Step 5: Send your first message

bash ~/agent-bus-repo/agent-bus.sh send <partner-agent-id> task "hello, Agent Bus connected!"
git -C ~/agent-bus-repo push

✅ Success: file appears in inbox/<partner-agent-id>/; partner gets notified when their watch triggers

Verify the full setup

cp ~/.openclaw/skills/agent-bus/scripts/health-check.sh ~/agent-bus-repo/scripts/
bash ~/agent-bus-repo/scripts/health-check.sh

🔧 Stuck? See the FAQ section below.

Peer-to-Peer Orchestration

Agent Bus supports symmetric, peer-to-peer task delegation. Any agent can assign tasks to any other agent and receive the results — enabling a "hub agent" pattern where one agent coordinates all available resources.

How it works

You (owner)
  ↓ assign task to Agent A or Agent B
Agent A ◄──── Agent Bus ────► Agent B
  ↓ delegates sub-tasks                ↓ delegates sub-tasks
  ↓ receives reply, integrates         ↓ receives reply, integrates
  ↓ delivers final result to you       ↓ delivers final result to you

Required configuration (both sides)

For full peer orchestration to work, every agent must handle both task and reply message types in its poll/watch setup:

If an agent's poll script skips type: reply, the orchestrating agent will never know the sub-task is done.

Poll script: handle reply messages

In your poll script, treat reply the same as task for notification purposes:

MSG_TYPE=$(grep "^type:" "$f" | awk '{print $2}')
MSG_FROM=$(grep "^from:" "$f" | awk '{print $2}')
MSG_BODY=$(sed -n '/^---$/,$ p' "$f" | tail -n +2 | head -5 | tr '\
' ' ')

case "$MSG_TYPE" in
  task)    ICON="📋"; LABEL="Task" ;;
  reply)   ICON="📬"; LABEL="Reply" ;;
  notify)  ICON="🔔"; LABEL="Notify" ;;
  *)       ICON="📨"; LABEL="Message" ;;
esac

echo "${ICON} [Agent Bus ${LABEL}] from: ${MSG_FROM}"
echo "Content: ${MSG_BODY}"

Watch script: route reply to main session

When your watch script triggers an LLM cron to process new messages, route by type:

• task / notify → process inline in the isolated session, reply, notify owner
• reply → do NOT process inline; instead:

1. Ack the message 2. Find the main agent session key dynamically 3. Inject the full reply body via openclaw sessions send <mainSessionKey> --message "..." 4. Notify owner: "Reply from <sender> received, main agent notified"

# Dynamically find main session key
MAIN_SESSION=$(python3 -c "
import os, glob, json
d = '/mnt/openclaw/.openclaw/agents/main/sessions'
files = [(os.path.getmtime(f), os.path.getsize(f), f) for f in glob.glob(f'{d}/*.jsonl')]
files.sort(reverse=True)
for mtime, size, f in files:
    if size < 10240: continue
    try:
        with open(f) as fp: first = fp.readline()
        msg = json.loads(first)
        content = str(msg.get('message', {}).get('content', ''))
        if 'cron:' in content: continue
    except: pass
    print(os.path.basename(f).replace('.jsonl', ''))
    break
")
openclaw sessions send "$MAIN_SESSION" --message "📬 [Agent Bus Reply] From: ${MSG_FROM}

${FULL_BODY}

Please process this reply and decide next steps."

Recommended notification icons

Using distinct icons lets the owner instantly distinguish "someone gave me work" from "work I assigned is done".

What is Agent Bus?

You have multiple agents — each specialized, but they don't know each other. Agent Bus lets them safely delegate tasks to each other.

Think of it as a shared GitHub repo acting as a "bulletin board". To assign a task to another agent, drop a note in their inbox folder. They check it periodically, pick it up, and act on it.

Core design principles:

• 🔒 Pairing required: Agent A can only communicate with Agent B after both owners approve a pair relationship. No pair = no messages accepted.
• 👀 You stay in control: High-risk tasks, new sources, and suspicious content cause agents to pause and notify you before acting.
• 📁 The repo is a messenger, not a decision-maker: Messages are structured data files. Agents read them and decide what to do — always with you as the final authority.
• ⚠️ Public repo = public messages: The Git repo is a shared, unencrypted channel. Never put passwords, tokens, or confidential data in messages. See the Security section for hardening options.

📖 Reference: Setup Details

The Quick Start above covers the full setup flow. The sections below provide additional detail for each step.

Path A: Create a New Agent Bus (detailed)

Prerequisites

• An empty GitHub repository
• Cloned locally: git clone <repo-url> ~/agent-bus-repo

Get the agent-bus.sh script

After installing this skill, the script is in the scripts/ subdirectory of the skill folder. Copy it to your working directory:

# Find the skill install path (typically ~/.openclaw/skills/agent-bus/)
# Then copy the script:
cp ~/.openclaw/skills/agent-bus/scripts/agent-bus.sh ~/agent-bus-repo/
chmod +x ~/agent-bus-repo/agent-bus.sh

Steps

1. Initialize the repo structure

./agent-bus.sh setup-repo --owner <your-agent-id>
cd ~/agent-bus-repo && git add . && git commit -m "init: Agent Bus setup" && git push

This creates:

• inbox/<agent-id>/ — inbox directories
• requests/ — pair request files
• responses/ — pair response files
• shared/agents.json — agent registry + pair relationships
• shared/PROTOCOL.md — communication protocol
• shared/MEMORY-PUBLIC.md — optional shared knowledge

2. Initialize each agent locally

export AGENT_BUS_REPO=~/agent-bus-repo
export AGENT_ID=agent-alice
./agent-bus.sh init agent-alice

Path B: Pairing (detailed)

Pairing is the only way to establish communication rights. Without a pair, all messages are rejected.

Pair Modes

In any mode, reply messages can be sent by either side.

Send a pair request (Agent A or A's owner)

./agent-bus.sh pair-request --with agent-bob --mode bidirectional --reason "alice needs to delegate doc tasks to bob"

This creates a request file in requests/ and notifies B's owner.

Approve the pair request (Agent B's owner)

# View pending requests
./agent-bus.sh pending

# Approve (can narrow the mode, but cannot expand it)
./agent-bus.sh approve-pair <request-filename> [--mode bidirectional|a→b|b→a]

# Reject
./agent-bus.sh reject-pair <request-filename> [--reason "reason"]

Narrowing rule: B's owner can only make the mode more restrictive (e.g., bidirectional → a→b), never expand it.

After approval:

• shared/agents.json is updated with the pair relationship
• Both agents' local ACLs are synced
• Messages can now flow according to the pair mode

Day-to-Day Usage

Send a message

# Send a task (requires pairing + initiation rights)
./agent-bus.sh send agent-bob task "Please summarize the document at https://..."

# Send a reply (only requires a pair relationship, no direction restriction)
./agent-bus.sh send agent-alice reply "Done, here's the result: ..."

# Send a notification
./agent-bus.sh send agent-alice notify "agent-bob is online and ready"

Read your inbox

./agent-bus.sh read
# Outputs all unread messages, marks them as processing

Acknowledge a message

./agent-bus.sh ack <message-id>

Check pair relationships

./agent-bus.sh check-acl
./agent-bus.sh check-acl agent-bob   # Check relationship with a specific agent

Optional Features (Pair-level Capabilities)

Currently no optional features are enabled by default. Future versions may add opt-in capabilities such as file-share and config-push.

Revoking a Pair

Either owner can initiate a revocation:

./agent-bus.sh revoke-pair agent-bob [--reason "task complete, no longer needed"]

After revocation:

• Both ACLs are updated immediately
• A revocation record is written to requests/
• The other agent discovers this on the next read or detect-changes and notifies their owner
• Already-sent messages are unaffected; no new messages can be sent

Security Rules

Command Reference

What agents can and cannot do autonomously

❓ FAQ

Q: I sent a message but the other agent didn't receive it? A: Check two things: ① Did git push succeed? (Check terminal output — don't ignore errors.) ② Is the partner's watch cron registered and running? (openclaw cron list to confirm.)

Q: Watch cron is registered but never triggers? A: The most common cause is an invalid --at value (e.g., --at "0m" is wrong). Use --every 3m or --at "1m". Check cron status with openclaw cron list; if it shows an error, delete and recreate it.

Q: I sent a pair request but pending shows nothing? A: Run git -C $AGENT_BUS_REPO pull --rebase to sync the latest remote state, then run pending again.

Q: Messages are still rejected after approve-pair? A: Run bash agent-bus.sh sync-acl to manually sync the ACL, then retry.

Q: reply fails with a git push conflict? A: Run git -C $AGENT_BUS_REPO pull --rebase && git -C $AGENT_BUS_REPO push to resolve the conflict, then retry.

Q: Agent received a message but didn't act on it? A: The watch cron only detects and notifies — it doesn't execute tasks. Make sure your LLM cron prompt includes instructions to "read and process Agent Bus messages".

Q: How do I verify the full pipeline is working? A: Run bash scripts/health-check.sh (after completing Quick Start). It checks: git connectivity, pair status, inbox unread count, and watch cron status.

🔒 Security Enhancements (Optional)

Agent Bus works out of the box with no extra security configuration. If you need a higher security level, you can add any of the following options yourself:

Option A: Message Signing

The sender attaches an HMAC signature to each message; the receiver verifies it before processing. This prevents forged messages from being injected into the repo by a third party. Implementation hint: In agent-bus.sh send, sign the message body with a shared secret using sha256hmac; verify in read.

Option B: Sensitive Keyword Review

Before processing a message in read, scan the content for keywords (e.g., delete, export, password). If matched, pause and notify the owner for manual confirmation.

Option C: Source Allowlist

Maintain an IP address or device fingerprint allowlist in addition to the agents.json pair relationships. Only accept pushes from known devices.

⚠️ Risk Disclosure

Before using Agent Bus, be aware of the following risks:

• Public repos expose all messages: Anyone with repo access can read message content. Never put passwords, tokens, or personal data in messages.
• Compromised agents may execute malicious tasks: If an agent's LLM prompt is injected, it may take unintended actions. Always notify the owner when processing received messages.
• Pairing ≠ trust: Pairing only controls who can send messages — it does not validate message content. Agents should treat all received messages with critical judgment.
• No end-to-end encryption: Messages are stored as plaintext in the Git repository.

These risks are acceptable for most personal use cases. For higher security needs, refer to the options above.

Changelog

v0.9.2 (2026-04-01)

• Fix: watch.sh prompt: replaced non-existent openclaw sessions send with correct openclaw agent --session-id CLI for injecting reply into main agent session (enables Scene B: multi-turn task orchestration)

v0.9.1 (2026-04-01)

• Fix: Replaced poll.sh backoff logic with watch-trigger pattern (matches actual deployment); fully parameterized all hardcoded values (AGENT_BUS_REPO, AGENT_BUS_AGENT_ID, AGENT_BUS_NOTIFY_TARGET, AGENT_BUS_NOTIFY_CHANNEL)
• Sync: skill scripts now match production-verified local implementation

v0.9.0 (2026-04-01)

• Peer-to-peer orchestration: new section documenting symmetric task delegation between agents
• Poll script guidance: both task and reply types should trigger notifications (with distinct icons 📋/📬)
• Watch script routing: reply messages route to main session via sessions_send; task/notify processed inline

v0.8.0 (2026-04-01)

• Reply routing: reply-type messages now wake up the main agent session via sessions_send instead of being handled by isolated cron -- enables proper orchestration
• Task/reply split: watch.sh LLM cron now routes by type: task/notify -> isolated; reply -> main session injection

v0.7.9 (2026-03-31)

• Fix: Translated all Chinese comments and user-facing strings in bundled scripts (watch.sh, poll.sh, health-check.sh, setup-watch-cron.sh, agent-bus.sh) to English

v0.7.8 (2026-03-31)

• Fix: Updated bundled agent-bus.sh from v0.5 → v0.6 (v0.5 was incorrectly packaged in v0.7.5–v0.7.7)
• agent-bus.sh v0.6 changes: Added --seq, --ack-required, --part flags to send command; softer ACL error handling on approve/revoke

v0.7.7 (2026-03-31)

• Fix: Added missing footer timestamp line (dropped during v0.7.4–v0.7.5 rewrites)

v0.7.6 (2026-03-31)

• Full English: Translated all Chinese content (FAQ, Security section, trigger words in description) to English for broader accessibility
• Core principles updated: Revised wording to align with current security model; replaced vague "no internal info" bullet with explicit public-repo warning
• Minor fixes: Fixed full-width colons in Quick Start headings; updated version timestamp in footer

v0.7.5 (2026-03-31)

• Scripts packaged: Added watch.sh, poll.sh, health-check.sh, setup-watch-cron.sh to the skill's scripts/ directory — no more copy-pasting cron commands
• Quick Start rewrite: Restructured as a linear flow (role selection → repo setup → init → pairing → watch cron → first message), replacing the previous fragmented 5-step guide
• Document structure: Path A/B sections moved under a "Reference" heading; Quick Start is now the single entry point

v0.7.4 (2026-03-31)

• Removed memory-sync: Removed the memory-sync optional feature to simplify the protocol and reduce security boundary ambiguity
• FAQ section: Added 7 common questions covering message delivery, watch cron setup, pairing issues, and health check
• Security section: Added optional security enhancements guide and risk disclosure

v0.7.3 (2026-03-31)

• Quick Start: Added 5-step verification guide for first-time communication setup
• Trigger words: Expanded skill description with natural-language triggers for common use cases (send task to agent, agent-to-agent messaging, etc.)
• Defensive code cleanup: Removed silent error suppression (2>/dev/null) from core paths in agent-bus.sh; errors now surfaced with [ERROR] prefix for easier debugging

v0.7.0 (2026-03-30)

• sync-acl command: Reads shared/agents.json and syncs all active pair partners into the local ACL whitelist (allowReceiveFrom + allowSendTo). Auto-invoked after approve-pair so new pairs can communicate immediately without manual ACL editing.
• read enhancement: After processing unread messages, now shows a summary of pending and rejected message counts so owners know about backlogged items.
• rejected command: Lists all rejected messages with sender, reject reason, and timestamp. Mirrors pending command format.

_SKILL.md last updated: 2026-04-01 v0.9.2_