Subagent Driven Development

Subagent-Driven Development

Iron Law: Complex features require specialized expertise, coordinated execution, and multi-stage verification.

When to Use

The Process

1. DECOMPOSE  -> Break feature into subtasks
2. SELECT     -> Assign agents to subtasks
3. SEQUENCE   -> Define execution order
4. DELEGATE   -> Invoke agents with full context
5. AGGREGATE  -> Collect and integrate results
6. VERIFY     -> Final quality check

Step 1: Decompose

Break the feature into discrete, testable subtasks:

## Feature: User Invitation System

### Subtasks:
1. Database schema for invitations (supabase-specialist)
2. RLS policies for invitation access (security-auditor)
3. API endpoints for invite CRUD (backend-engineer)
4. Email notification service (backend-engineer)
5. Invitation UI components (frontend-specialist)
6. Integration tests (quality-engineer)

Requirements:

• Each subtask is independently testable
• Clear inputs and outputs defined
• Dependencies between subtasks identified
• Estimated scope per subtask (small/medium/large)

Step 2: Agent Selection

Selection Criteria:

• Match agent expertise to subtask domain
• Consider agent tool permissions
• Prefer specialists over generalists
• Use orchestrator for coordination-heavy tasks

Step 3: Sequence Planning

Dependency Chain Example:

[1] Database Schema (supabase-specialist)
    ↓
[2] Security Policies (security-auditor) - depends on [1]
    ↓
[3] Backend APIs (backend-engineer) - depends on [1], [2]
    ↓
[4] Email Service (backend-engineer) - parallel with [3]
    ↓
[5] Frontend UI (frontend-specialist) - depends on [3]
    ↓
[6] Integration Tests (quality-engineer) - depends on [3], [4], [5]

Sequencing Rules:

• Database before API
• Security review before implementation
• Backend before frontend (for API contracts)
• Implementation before testing
• Parallelize independent subtasks

Step 4: Delegation Protocol

Task Tool Invocation Format:

Task: @[agent-name]

## Context
[Background information about the feature]

## Your Subtask
[Specific work to complete]

## Required Skills
Load skills: [skill-1], [skill-2]

## Inputs
[Data/files from previous subtasks]

## Expected Output
[Specific deliverables expected]

## Quality Criteria
- [ ] Criterion 1
- [ ] Criterion 2

## Handoff Instructions
[What to prepare for next agent]

Delegation Rules:

• Provide FULL context (don't assume agent knows background)
• Include all relevant file paths
• Specify exact skills to load
• Define clear success criteria
• Include handoff instructions for next agent

Step 5: Result Aggregation

After each agent completes:

## Subtask [N] Complete

### Agent: [agent-name]
### Status: SUCCESS | PARTIAL | BLOCKED

### Deliverables:
- [x] File created: /path/to/file.ts
- [x] Migration added: /path/to/migration.sql
- [ ] Documentation updated (pending)

### Key Decisions:
- Decision 1: Chose approach A because [reason]
- Decision 2: Used pattern B for [reason]

### Issues/Blockers:
- None | [Description of issue]

### Handoff Data:
[Information needed by subsequent agents]

Aggregation Rules:

• Document all deliverables with paths
• Record key architectural decisions
• Track any blockers or issues
• Prepare handoff data for next agent
• Update session file with progress

Step 6: Final Verification

## Feature Verification Checklist

### Functional
- [ ] All subtasks completed successfully
- [ ] Integration points working
- [ ] User flows tested end-to-end

### Quality
- [ ] Tests pass (unit, integration, E2E)
- [ ] No TypeScript errors
- [ ] Lint rules satisfied
- [ ] Security review passed

### Documentation
- [ ] API documentation updated
- [ ] Architecture decisions recorded
- [ ] README updated if needed

Handoff Protocols

Between Agents:

Agent A → Agent B
├── Summary of completed work
├── Files created/modified (full paths)
├── Interfaces/types defined
├── Environment variables needed
└── Known constraints or decisions

Back to Orchestrator:

Agent → Orchestrator
├── Completion status
├── All deliverables listed
├── Issues encountered
├── Recommendations for next steps
└── Estimated impact on timeline

Common Patterns

Pattern: Security-First Feature

1. Research → Requirements
2. Security Audit → Risk assessment
3. Database → Schema with RLS
4. Security Review → Policy verification
5. Backend → Secure implementation
6. Frontend → Secure UI
7. Quality → Security tests

Pattern: Performance-Critical Feature

1. Research → Best practices
2. Performance → Baseline metrics
3. Database → Optimized schema
4. Backend → Efficient queries
5. Frontend → Optimized rendering
6. Performance → Verification
7. Quality → Load tests

Red Flags

• Skipping decomposition -> Unclear scope, missed dependencies
• Single agent for complex feature -> Missing expertise
• No handoff data -> Lost context between agents
• Skipping verification -> Integration failures
• Parallel when sequential needed -> Race conditions

Decision Criteria

Integration

Pairs with: dispatching-parallel-agents (for independent subtasks), verification (final check), session (complex workflows)