workflow-orchestration（工作流程编排）

Workflow Orchestration

Orchestrate complex workflows by breaking them into parallel sub-tasks executed by specialized subagents, dramatically reducing completion time while maintaining quality.

When to Use

• Complex multi-step tasks that can be decomposed
• Tasks requiring multiple domain expertise areas
• Time-sensitive operations needing parallel execution
• Work requiring research, implementation, testing, and documentation simultaneously
• Any task where "throwing more compute at it" via parallel agents makes sense

Core Strategy

1. Task Decomposition

Break complex tasks into independent, parallel-executable subtasks:

Example: Authentication System Implementation

Sequential (5+ hours):
Research → Design → Code → Test → Document → Review

Parallel (1 hour):
├── research-specialist: OAuth 2.1 patterns research
├── auth-specialist: Core authentication implementation
├── test-specialist: Test suite creation
├── documentation-writer: API documentation
└── code-auditor: Security review

2. Agent Selection

Match subtasks to specialized agents:

3. Clear Boundaries

Define exactly what each agent should accomplish:

## Agent Task Specification

**Agent**: research-specialist
**Task**: Research OAuth 2.1 PKCE flow
**Deliverable**:
- Summary of PKCE benefits
- Code examples in TypeScript
- Security considerations
**Constraints**: Focus on browser-based flows only

Implementation Pattern

Step 1: Analyze and Decompose

Original Task: "Build a complete SaaS billing system"

Decomposition:
1. Database schema design (db-specialist)
2. Stripe integration (payments-specialist)
3. Billing UI components (design-specialist)
4. Webhook handling (integration-expert)
5. Testing (test-specialist)
6. Documentation (documentation-writer)

Step 2: Launch Parallel Agents

Use the Task tool to launch agents simultaneously:

Task 1: db-specialist - Design database schema
Task 2: payments-specialist - Set up Stripe integration
Task 3: design-specialist - Create billing UI components
Task 4: integration-expert - Implement webhooks
Task 5: test-specialist - Write test suite
Task 6: documentation-writer - Document the API

Step 3: Coordinate Results

Collect all agent outputs and integrate:

## Integration Checklist

- [ ] Database schema approved
- [ ] Stripe integration tested
- [ ] UI components match design system
- [ ] Webhooks properly secured
- [ ] Tests passing
- [ ] Documentation complete

Best Practices

DO

• Give each agent a single, focused task
• Provide detailed context in task descriptions
• Set clear deliverables and constraints
• Use 3-5 agents for most complex tasks
• Plan integration points before starting

DON'T

• Overload one agent with multiple responsibilities
• Launch agents without clear deliverables
• Forget to plan how results will be integrated
• Use parallel execution for simple sequential tasks
• Skip the decomposition analysis phase

Example Workflows

Feature Implementation

Task: "Add user profiles with avatars"

Parallel:
├── design-specialist: Avatar upload UI
├── db-specialist: User profile schema
├── integration-expert: Image storage (S3/Cloudinary)
└── test-specialist: Profile update tests

Security Audit

Task: "Audit authentication system"

Parallel:
├── code-auditor: Vulnerability scan
├── auth-specialist: Auth flow review
├── test-specialist: Penetration tests
└── documentation-writer: Security report

Documentation Sprint

Task: "Document entire API"

Parallel:
├── documentation-writer: Endpoint docs
├── code-auditor: Code example verification
└── test-specialist: Test case documentation

Error Handling

When a subagent fails:

1. Retry: Simple failures (timeouts, temporary errors)
2. Reassign: Task misalignment → different specialist
3. Decompose further: Task too large → break into smaller pieces
4. Sequential fallback: Dependencies discovered → reorder tasks

Performance Tips

• Agent Reuse: Keep successful agents for related follow-up tasks
• Context Preservation: Pass key findings between related agents
• Result Caching: Store agent outputs for reuse in future tasks
• Progress Tracking: Monitor all agents' progress simultaneously

Integration with Other Skills

• reinforce-skills: Persist workflow patterns to CLAUDE.md
• critique: Review parallel agent results for quality
• confess: Audit if parallel approach was optimal

References

• Parallel Agents Strategy
• Agent Collaboration Protocol
• Completion Reporting