subagent-brief

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":[]}}}

Sub-agent brief (OpenClaw)

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.

Before you spawn

1. Self-contained brief — Sub-agents receive 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.
2. Right context mode (native sub-agents):
   • 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.
3. Tool profile — 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.

Writing the task string

Include:

• Goal (one sentence) and done criteria (“return X”, “files changed under Y”).
• Inputs (paths, URLs, quoted constraints) — don’t assume the child saw the parent chat unless you used fork.
• Forbidden actions if relevant (e.g. “do not push to remote”).
• Output shape: bullets, patch summary, or structured JSON if orchestration needs it.

After spawn — do not poll in a tight loop

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.

Models and cost

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.

Nesting (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.

Silent / skip behaviors

• Reply exactly ANNOUNCE_SKIP in the child if nothing should be posted back.
• NO_REPLY / no_reply suppresses announce noise when appropriate.

Gateway restart

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.

Related

• Sub-agents
• Multi-agent routing
• Multi-agent sandbox and tools