Menu
Spawn effective OpenClaw sub-agents — clear briefs, isolated vs fork context, no polling loops, cheaper models for workers.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
End-to-end software work—read codebase, implement, test, and summarize changes like a senior engineer.
UX/UI feedback and copy tuned for clarity, accessibility, and consistency—without redesigning everything unprompted.
Lawful, ethical open-source intelligence on public data only—no stalking, no bypassing access controls, no invasive targeting.
Rigorous public research using web search, fetch, and source triangulation. Use for technical docs, competitors, and fact-checking.
| name | subagent-brief |
| description | Spawn effective OpenClaw sub-agents — clear briefs, isolated vs fork context, no polling loops, cheaper models for workers. |
| metadata | {"openclaw":{"requires":{"config":[]}}} |
When to use: Delegating parallel research, long-running tasks, or heavy tool use without blocking the main chat — via sessions_spawn, /subagents spawn, or equivalent.
AGENTS.md + TOOLS.md only (not SOUL.md, USER.md, HEARTBEAT.md). Put must-have constraints in the task text or ensure they live in AGENTS.md / TOOLS.md.isolated (default) — Fresh child transcript. Use for independent research, implementation from a written spec, or anything you can fully specify in the prompt. Lower token cost.fork — Branches the parent transcript into the child. Use only when the child needs prior conversation or tool results not practical to restate. Use rarely.sessions_spawn is available on coding / full profiles by default. On messaging, add tools.alsoAllow for spawn tools or switch profile. Confirm with /tools in-session.Include:
fork.Completion is push-based: the child announces back when done. Do not repeatedly call /subagents list, sessions_list, or sessions_history just to wait. Check status only for debugging or user-requested intervention.
Use sessions_history for bounded recall when synthesizing child output (sanitized). Inspect on-disk transcript path only when you need raw detail.
Each sub-agent has its own context and token usage. Prefer a cheaper / faster model for children via agents.defaults.subagents.model or spawn overrides when the task is mechanical.
maxSpawnDepth ≥ 2)Depth-1 can orchestrate depth-2 workers; depth-2 cannot spawn further. Stop/kill behavior cascades to children. Keep fan-out within maxChildrenPerAgent.
ANNOUNCE_SKIP in the child if nothing should be posted back.NO_REPLY / no_reply suppresses announce noise when appropriate.Announce delivery is best-effort; after a gateway restart, pending completion handoff may be lost. Don’t assume the child finished — verify if it matters.