openclaw-subagents（OpenClaw subagents 搜索）

OpenClaw Subagents Configuration Skill

Overview

Configure and manage OpenClaw subagents — background agent runs spawned from a main agent that operate in isolated sessions and report results back upon completion. This skill covers everything from a single subagent spawn to full multi-agent squad orchestration.

References

Load these files as needed during configuration:

• references/config-reference.md — Full openclaw.json schema for subagents (all fields, defaults, per-agent overrides)
• references/sessions-spawn-tool.md — sessions_spawn tool parameters, slash commands, and behavior details
• references/agent-files.md — SOUL.md, AGENTS.md, HEARTBEAT.md, and memory file templates
• references/multi-agent-architecture.md — Depth levels, tool access rules, announce chain, Mission Control integration

Step 1: Determine Setup Type

Ask or infer from context which pattern the user needs:

Step 2: Configure openclaw.json

Config lives at ~/.openclaw/openclaw.json. Load references/config-reference.md for the full schema.

Minimal subagents config:

{
  agents: {
    defaults: {
      subagents: {
        model: "anthropic/claude-haiku-4-5",  // use cheaper model for subagents
        maxSpawnDepth: 1,                      // set to 2 for orchestrator pattern
        maxChildrenPerAgent: 5,               // max active children per session (1-20)
        maxConcurrent: 8,                     // global concurrency limit
        runTimeoutSeconds: 900,               // abort after N seconds (0 = no limit)
        archiveAfterMinutes: 60               // auto-delete transcript after N minutes
      }
    },
    list: [
      {
        id: "main",
        workspace: "~/.openclaw/workspace-main"
      }
    ]
  }
}

For per-agent model overrides and tool restrictions, see references/config-reference.md.

Step 3: Create Agent Identity Files

Each agent workspace at ~/.openclaw/workspace-<agentId>/ needs identity files. Load references/agent-files.md for full templates.

Step 4: Set Up Cron Heartbeats

Each agent wakes every 15 minutes via cron. Stagger agents by 2 minutes each:

openclaw cron add \
  --name "agent-heartbeat" \
  --cron "0,15,30,45 * * * *" \
  --session "isolated" \
  --message "You are <AgentName>, the <Role>. Read WORKING.md. Check Mission Control for @mentions and assigned tasks. If work exists, do it and update WORKING.md. If nothing to do, reply HEARTBEAT_OK."

Suggested stagger schedule for a squad:

:00 Agent 1 (e.g. Pepper)
:02 Agent 2 (e.g. Shuri)
:04 Agent 3 (e.g. Friday)
:06 Agent 4 (e.g. Loki)
:08 Agent 5 (e.g. Vision)

Step 5: Spawning Subagents

Subagents are spawned non-blocking — the tool returns a run ID immediately and the result is announced on completion.

From within an agent (agent calls sessions_spawn tool with these params):

task: "Research competitor pricing for task #42 and post findings as a comment"
agentId: "researcher"          // optional: target specific agent
label: "competitor-research"   // optional: human-readable name
model: "claude-haiku-4-5"     // optional: override model
thinking: "none"               // none | basic | deep
runTimeoutSeconds: 600         // optional: abort after N seconds
cleanup: "delete"              // delete = archive immediately after announce
mode: "run"                    // run = one-shot (default), session = persistent

Manually via slash command:

/subagents spawn researcher "Research competitor pricing for task #42"

Load references/sessions-spawn-tool.md for full parameter details.

Step 6: Verify and Monitor

openclaw agents list --bindings   # confirm agent routing
openclaw gateway start            # start the gateway daemon

In chat:

/subagents list                   # view active runs
/subagents info <id>              # status, timestamps, session key
/subagents log <id> [limit]       # execution logs
/subagents steer <id> <message>   # redirect mid-run
/subagents kill <id|all>          # stop subagent + cascade to children

Tool Access by Depth (Critical)

Blocked for all subagents by default: sessions_list, sessions_history, sessions_send, sessions_spawn

Common Pitfalls

• Never share agentDir across agents — auth profiles are per-agent and must remain isolated
• SOUL.md is not injected into subagents — put operational instructions in AGENTS.md instead
• Memory is files only — anything not written to disk is lost when a session ends
• Always stagger heartbeat crons — running all agents at :00 causes resource spikes
• Use cheaper models for heartbeats; reserve powerful models for creative or reasoning-heavy tasks
• Announce is best-effort — if the gateway restarts mid-run, the announce may be lost; design tasks to be resumable
• Max nesting depth is 5, but 2 is the recommended maximum for practical use