| name | multi-agent-bootstrap |
| description | Bootstrap and migrate an OpenClaw workspace from single-agent to beginner-friendly multi-agent mode with visible routing. Use when user asks to create/adapt/sync multi-agent setup, add role definitions, enforce routing line output, or quickly replicate a proven multi-agent architecture. |
| requires | ["file_write"] |
Multi-Agent Bootstrap
Create or update a workspace to a stable multi-agent baseline for beginners.
Workflow
- Read current files first:
AGENTS.md, USER.md, MULTI_AGENTS.md (if exists), MEMORY.md (if main session).
- If
MULTI_AGENTS.md is missing, create a minimal version with CTRL contract + 4 core roles.
- Ensure role set includes at least:
LIFE/WORK/LEARN/ENGINEER.
- Ensure routing visibility rule is explicit in both
AGENTS.md and MULTI_AGENTS.md:
Routing: User -> CTRL -> [ROLE] -> CTRL -> User
- Parallel:
Routing: User -> CTRL -> ([ROLE_A] || [ROLE_B]) -> CTRL -> User
- Apply beginner safety defaults:
- default single specialist routing
- max parallel specialists = 2
- CTRL is final merger only
- If user asks “固化/记住”, persist in both:
MULTI_AGENTS.md (source of truth)
MEMORY.md (preference continuity)
- Run
tools/check_skill_pack.sh to validate required fields/rules (if available).
- If validation passes, summarize changes and recommend 3 smoke tests.
Editing Rules
- Prefer surgical edits over full rewrites.
- Keep role labels business-facing; avoid generic tags like PLAN.
- Do not silently remove existing user-specific roles.
- On conflicts: safety/boundary rules in
AGENTS.md win.
Required Outputs
- A concise change summary.
- Explicit routing line in every user-facing reply.
- A short “what to test next” list.