agent-hive（Agent 蜂巢）

Last used: 2026-03-24 Memory references: 2 Status: Active

Agent Hive

Create and manage multi-agent teams with shared memory, budget governance, and configurable spawn permissions.

Architecture

Main Workspace (~/.openclaw/workspace/)
├── MEMORY.md, TOOLS.md, USER.md, IDENTITY.md   ← shared brain (real files)
├── agents/<name>/                                ← per-agent outbox, inbox, budget
├── agents/governance/                            ← governance rules + audit log
├── memory/, projects/, scripts/, skills/         ← shared resources
└── content/                                      ← shared content

Agent Workspace (~/.openclaw/workspace-<id>/)
├── SOUL.md              ← unique personality (local file)
├── AGENTS.md            ← unique instructions (local file)
├── HEARTBEAT.md         ← agent-specific heartbeat (local file)
├── .openclaw/           ← agent-specific config dir
└── everything else      ← SYMLINKS to main workspace

Adding a New Agent

Step 1: Create workspace with symlinks

AGENT_ID="<id>"
MAIN="$HOME/.openclaw/workspace"
WS="$HOME/.openclaw/workspace-$AGENT_ID"

mkdir -p "$WS/.openclaw"

# Symlink shared brain
for f in .learnings IDENTITY.md MEMORY.md ROADMAP.md TOOLS.md USER.md; do
  ln -sf "../workspace/$f" "$WS/$f"
done
for d in agents content memory projects scripts skills; do
  ln -sf "../workspace/$d" "$WS/$d"
done

Step 2: Create unique files

Create $WS/SOUL.md — the agent's personality. This is theirs alone.

Create $WS/AGENTS.md — agent-specific instructions. Must include:

• Session startup checklist (read SOUL.md, check BUDGET.json, check INBOX.md)
• Communication rules (INBOX/OUTBOX pattern)
• What the agent does and doesn't do
• Budget rules section (see references/budget-rules.md)

Create $WS/HEARTBEAT.md — minimal heartbeat config.

Step 3: Create agent directory in shared workspace

mkdir -p "$MAIN/agents/$AGENT_ID"
touch "$MAIN/agents/$AGENT_ID/INBOX.md"
touch "$MAIN/agents/$AGENT_ID/OUTBOX.md"

Step 4: Create budget file

Copy scripts/create_budget.sh output or create manually:

{
  "daily_limit_output_tokens": 50000,
  "today": "YYYY-MM-DD",
  "used_output_tokens": 0,
  "spawns": [],
  "status": "active",
  "warnings": [],
  "consecutive_overbudget_days": 0
}

Save to $MAIN/agents/$AGENT_ID/BUDGET.json.

Step 5: Register in openclaw.json

Add to agents.list[]:

{
  "id": "<id>",
  "name": "<Name>",
  "workspace": "/absolute/path/to/workspace-<id>",
  "identity": { "name": "<Name>", "emoji": "<emoji>" },
  "model": {
    "primary": "anthropic/claude-sonnet-4-6",
    "fallbacks": ["<fallback-model-1>", "<fallback-model-2>"]
  },
  "subagents": { "allowAgents": ["<peer1>", "<peer2>"] }
}

Update existing agents' allowAgents to include the new agent.

Step 6: Validate and restart

python3 -c "import json; json.load(open('$HOME/.openclaw/openclaw.json')); print('JSON valid')"
openclaw doctor  # check for schema errors
launchctl stop ai.openclaw.gateway && sleep 2 && launchctl start ai.openclaw.gateway

Wait 15 seconds before testing.

Spawn Permission Models

Hub-and-spoke (conservative)

Only the orchestrator (main) can spawn agents. Agents route requests through OUTBOX.

"subagents": { "allowAgents": [] }  // for all non-main agents

Full mesh with budget (recommended)

All agents can spawn peers. Budget governance prevents abuse.

// Each agent can spawn all peers (not themselves)
"subagents": { "allowAgents": ["<all other agent ids>"] }

Earned mesh (progressive trust)

Start hub-and-spoke. Promote to mesh after demonstrating budget responsibility. Demote back to hub-and-spoke after repeated overspend.

Budget Governance

See references/governance.md for the full framework.

Run the audit script on heartbeat:

python3 scripts/budget_audit.py

Thresholds: Green (<80%), Yellow (80-100%), Red (>100%), Emergency demotion (>200%). 3 consecutive overbudget days → automatic demotion (mesh privileges revoked).

Add step 4.5 to your HEARTBEAT.md — see references/heartbeat-snippet.md.

Committee Pattern

For project work, compose agent committees:

1. Identify which agents are needed (e.g., content + marketing + design review)
2. Spawn them in parallel with clear, scoped tasks
3. Each writes to their OUTBOX.md
4. Orchestrator collects results and synthesizes
5. Budget tracked per-agent across the project

Verification Checklist

After adding an agent, verify:

• [ ] workspace-<id>/ has SOUL.md + AGENTS.md as local files
• [ ] All other files are symlinks to main workspace
• [ ] agents/<id>/BUDGET.json exists with correct schema
• [ ] Agent appears in openclaw doctor output
• [ ] Gateway restarts without errors
• [ ] Agent can be spawned: sessions_spawn(agentId="<id>", task="Say hello")
• [ ] Budget audit includes the new agent: python3 scripts/budget_audit.py