Multi-Agent Workflows
Agents: @orchestrator @executor @planner · Complexity: Medium-High
Real-world workflow examples showing how multiple agents coordinate to deliver complete features.
Startup MVP Sprint (1 week)
Goal: Build a SaaS MVP with auth, dashboard, and billing.
Agents involved: @planner → @starter → @orchestrator → @builder × 3 → @tester → @reviewer
Workflow
Day 1: Planning & Setup
──────────────────────
User: /plan Build a SaaS MVP with user auth, dashboard, and Stripe billing
@planner analyzes requirements
→ Produces COMPLEX plan with 12 tasks
User: Create the project
@starter scaffolds project
→ Frontend (React/Vue) + Backend (Node.js) + Database (PostgreSQL)
→ Installs Specialist Agent agents and skills
Day 2-3: Parallel Build
────────────────────────
User: @orchestrator Execute the plan
@orchestrator decomposes into 3 parallel streams:
Stream 1: @builder (Auth)
├── User model + migration
├── JWT auth middleware
├── Login/Register endpoints
└── Auth composable/hook
Stream 2: @builder (Dashboard)
├── Dashboard layout
├── Stats service + adapter
├── Chart components
└── Dashboard page
Stream 3: @builder (Billing)
├── Stripe integration service
├── Subscription plans
├── Checkout flow
└── Webhook handler
Handoff: @orchestrator merges all streams
→ Runs integration tests
→ Creates checkpoint
Day 4: Testing & Review
───────────────────────
User: @tester Create test strategy for the MVP
@tester designs testing pyramid:
→ 70% unit tests (services, adapters)
→ 20% integration tests (API endpoints)
→ 10% E2E tests (auth flow, checkout flow)
User: @reviewer Review all changes
@reviewer runs unified 3-in-1 review:
→ Spec Compliance: PASS
→ Code Quality: B+ (minor improvements suggested)
→ Architecture Fit: PASS
Day 5: Polish & Deploy
──────────────────────
User: @devops Set up CI/CD pipeline
@devops creates:
→ GitHub Actions workflow
→ Docker Compose for local dev
→ Deployment configCost Estimate
@planner: ~3,000 tokens
@starter: ~8,000 tokens
@orchestrator: ~5,000 tokens
@builder × 3: ~45,000 tokens
@tester: ~12,000 tokens
@reviewer: ~8,000 tokens
@devops: ~6,000 tokens
─────────────────────────────
Total: ~87,000 tokens (~$1.30)Landing Page Sprint (1 day)
Goal: Design and build a complete landing page in one day.
Agents involved: @designer → @builder × 2 → @perf → @reviewer
Workflow
Morning: Design & Setup (9:00 - 11:00)
───────────────────────────────────────
User: @designer Create a design system for our landing page
@designer produces:
→ Color palette, typography, spacing tokens
→ Component inventory (Hero, Features, Pricing, CTA, Footer)
→ Responsive breakpoints
→ Accessibility requirements (WCAG 2.1 AA)
Midday: Parallel Build (11:00 - 15:00)
──────────────────────────────────────
User: @orchestrator Build the landing page sections in parallel
@orchestrator assigns:
@builder-1 (Above the fold):
├── Hero section with CTA
├── Social proof / logos
└── Feature highlights
@builder-2 (Below the fold):
├── Detailed features grid
├── Pricing table
├── FAQ accordion
└── Footer with links
Handoff: Merge + verify responsive behavior
Afternoon: Optimize & Ship (15:00 - 17:00)
──────────────────────────────────────────
User: @perf Optimize the landing page for Core Web Vitals
@perf analyzes:
→ Image optimization (WebP, lazy loading)
→ Bundle size (code splitting, tree shaking)
→ Font loading strategy (font-display: swap)
→ CLS fixes (explicit dimensions)
User: @reviewer Final review
@reviewer verdict: Approved
→ Performance: A (LCP < 2.5s)
→ Accessibility: A (WCAG 2.1 AA compliant)
→ SEO: Meta tags, structured data, sitemapBug Investigation & Fix (2 hours)
Goal: Investigate a production bug, find root cause, fix, and verify.
Agents involved: @doctor → @debugger → @builder → @tester → @reviewer
Workflow
Phase 1: Diagnosis (30 min)
───────────────────────────
User: Users report "Payment failed" errors intermittently
@doctor runs 4-phase diagnosis:
Phase 1 - Symptoms: Error logs show Stripe webhook 500s
Phase 2 - Hypotheses:
H1: Race condition in webhook handler
H2: Idempotency key not set
H3: Database transaction timeout
Phase 3 - Evidence:
→ Logs show duplicate webhook events
→ Handler processes same event twice
→ Second processing fails on unique constraint
Phase 4 - Root Cause: Missing idempotency check
Phase 2: Fix (30 min)
─────────────────────
User: @builder Fix the webhook handler
@builder implements:
→ Add idempotency key check
→ Add database transaction wrapping
→ Add structured logging for webhook events
→ Add retry logic with exponential backoff
Phase 3: Verify (30 min)
────────────────────────
User: @tester Write tests for the fix
@tester creates:
→ Unit test: idempotency check
→ Integration test: duplicate webhook handling
→ Edge case: concurrent webhook processing
User: /verify all
/verify output:
→ Tests: 47/47 passing (3 new)
→ TypeScript: 0 errors
→ Build: success
Phase 4: Review (30 min)
────────────────────────
User: @reviewer Review the fix
@reviewer verdict: Approved
→ Security: A (proper input validation)
→ Fix addresses root cause, not symptoms
→ Observability: structured logging addedMigration Project (1-2 weeks)
Goal: Migrate a legacy Vue 2 Options API app to Vue 3 Composition API.
Agents involved: @scout → @planner → @orchestrator → @migrator × N → @tester → @reviewer
Workflow
Week 1: Analysis & Planning
────────────────────────────
User: @scout Analyze the current codebase
@scout produces inventory:
→ 45 components (38 Options API, 7 already Composition)
→ 12 mixins (need conversion to composables)
→ 8 Vuex modules (need migration to Pinia)
→ 6 filters (need conversion to functions)
→ Dependency audit: 3 packages need Vue 3 alternatives
User: /plan Migrate the entire app to Vue 3 Composition API
@planner produces COMPLEX plan:
→ Phase 1: Infrastructure (Vue 3, Vite, Pinia)
→ Phase 2: Shared code (mixins → composables, filters)
→ Phase 3: Components (batch migration, 5/day)
→ Phase 4: Testing (update all tests)
→ Phase 5: Cleanup (remove Vue 2 compatibility)
Week 2: Execution
──────────────────
User: @orchestrator Execute the migration plan
@orchestrator coordinates:
Day 1: @migrator - Infrastructure migration
Day 2: @migrator × 2 - Mixins → Composables (parallel)
Day 3-4: @migrator × 3 - Components (batches of 5)
Day 5: @tester - Update all tests
Final: @reviewer - Full codebase review
Handoff template used at each phase boundary:
→ QA PASS / QA FAIL with evidence
→ Context transfer to next phase agentMulti-Domain Feature (3-5 days)
Goal: Add a referral system with frontend, backend, database, and notifications.
Agents involved: @analyst → @planner → @data → @api → @orchestrator → @builder × 2 → @tester → @security → @reviewer
Workflow
Day 1: Requirements & Design
─────────────────────────────
User: @analyst Define requirements for a referral system
@analyst produces spec:
→ User stories (referrer, referee)
→ Business rules (reward tiers, limits)
→ Acceptance criteria per story
→ Edge cases (self-referral, expired links)
User: @data Design the database schema
@data produces:
→ referral_codes table
→ referral_events table
→ reward_transactions table
→ Indexes and constraints
→ Migration script
User: @api Design the API endpoints
@api produces:
→ POST /api/referrals/generate
→ GET /api/referrals/:code
→ POST /api/referrals/redeem
→ GET /api/referrals/stats
→ OpenAPI spec
Day 2-3: Build
───────────────
User: @orchestrator Build the referral system
@orchestrator assigns parallel streams:
@builder-1 (Backend):
├── Referral service
├── Reward calculation
├── Email notifications
└── API endpoints
@builder-2 (Frontend):
├── Referral dashboard
├── Share/invite component
├── Reward history
└── Referral link handler
Day 4: Security & Testing
──────────────────────────
User: @security Audit the referral system
@security checks:
→ Rate limiting on code generation
→ Referral code entropy (not guessable)
→ Reward abuse prevention
→ Input validation on all endpoints
User: @tester Create comprehensive tests
@tester creates:
→ Unit: reward calculation, code generation
→ Integration: full referral flow
→ E2E: invite → signup → reward
Day 5: Review & Ship
─────────────────────
User: @reviewer Review the entire referral system
@reviewer unified 3-in-1:
→ Spec: All user stories implemented
→ Quality: A- (one minor improvement)
→ Architecture: Follows ARCHITECTURE.md
→ Verdict: ApprovedTips for Multi-Agent Workflows
- Always start with
/plan- Even if you think you know the scope - Use
@orchestratorfor 3+ agents - Don't manually coordinate - Checkpoints are mandatory - Every phase boundary gets a checkpoint
- Handoff templates - Use structured handoffs between agents
- Verify at each boundary - Run
/verifybefore moving to next phase - Cost tracking - Use
/estimatebefore expensive operations