安装
claude skill add --url github.com/openclaw/skills/tree/main/skills/amdf01-debug/sw-agent-team-orch文档
Trigger
Set up multi-agent teams with defined roles, task lifecycles, handoff protocols, and review workflows.
Trigger phrases: "multi-agent team", "agent orchestration", "set up agents", "task routing", "agent handoff", "agent coordination"
Process
- Define roles: What each agent specialises in
- Task lifecycle: inbox → spec → build → review → done
- Handoff protocol: How agents pass work between each other
- Quality gates: Review checkpoints before work moves forward
- Shared state: How agents share context and artifacts
Team Architecture Template
markdown
# Agent Team: [Name]
## Roles
### Manager Agent
- Routes incoming tasks to specialists
- Reviews completed work before delivery
- Escalates blocked tasks to human
- Model: [recommended model for this role]
### Specialist Agent: [Role Name]
- Handles: [task types]
- Outputs: [deliverable format]
- Quality bar: [minimum criteria]
- Model: [recommended model]
## Task Lifecycle
1. **Inbox**: New task arrives → Manager triages
2. **Assigned**: Manager routes to specialist with brief
3. **In Progress**: Specialist works, updates shared state
4. **Review**: Manager (or reviewer agent) checks output
5. **Revision**: If quality gate fails → back to specialist with notes
6. **Done**: Approved → delivered to requester
## Handoff Protocol
- Include: task description, context, acceptance criteria, deadline
- Never: assume context from previous tasks — always be explicit
- Format: structured JSON or markdown brief
## Quality Gates
- [ ] Output matches acceptance criteria
- [ ] No hallucinated data
- [ ] Formatting matches specification
- [ ] All links/references verified
- [ ] Spell-checked and proofread
Rules
- One task per agent at a time (focus > multitasking)
- Always include acceptance criteria in task briefs
- Shared state in files, not in agent memory (survives restarts)
- Model selection matters: use cheap models for bulk, expensive for judgment