Agent Architect

"I don't write code. I architect the architects who write the code."

A meta-agent skill for Claude Code that designs and orchestrates optimal AI agent swarms to build complex software projects.

Overview

Agent Architect transforms Claude Code from a single AI assistant into a software development firm. When you describe a project, it:

1. Analyzes your requirements
2. Designs the optimal team structure
3. Generates perfect prompts for each agent
4. Launches parallel execution
5. Monitors progress
6. Integrates all outputs
7. Delivers a polished final product

Installation

Copy the skill to your Claude Code skills directory:

# Copy entire skill folder
cp -r agent-architect ~/.claude/skills/

# Or create a symlink
ln -s /path/to/agent-architect ~/.claude/skills/agent-architect

Usage

Simply describe what you want to build:

You: Build me a social media scheduler that lets me queue posts
     for Twitter, LinkedIn, and Instagram, with a nice dashboard

Agent Architect: Analyzing... Designing swarm of 8 agents... Launching...

Trigger Phrases

The skill activates on:

• "Build me a..."
• "Create a full..."
• "I want to build..."
• "Make a complete..."
• "Ship a..."
• "Develop a..."
• "Architect a..."
• "Design and build..."
• "Use agent swarm to build..."
• "Parallel build..."

Directory Structure

agent-architect/
├── SKILL.md              # Main skill instructions
├── README.md             # This file
├── resources/
│   ├── config.yaml       # Configuration options
│   └── agent-roles.md    # Role reference guide
├── templates/
│   ├── agent-prompt.md   # Agent prompt template
│   ├── interfaces.md     # Interface contract template
│   └── swarm-status.json # Status tracking template
├── examples/
│   ├── social-scheduler-swarm.md  # Full example
│   └── chrome-extension-swarm.md  # Compact example
└── scripts/
    ├── init-workspace.sh    # Initialize workspace
    ├── check-status.sh      # Check swarm status
    └── integrate.sh         # Merge agent outputs

How It Works

Phase 1: Analysis

Parses your request to identify:

• Core product being built
• Required features
• Technology preferences
• Complexity level

Phase 2: Design

Creates optimal swarm architecture:

• Determines agent count (3-12)
• Assigns roles from templates
• Maps dependencies
• Creates execution waves

Phase 3: Generation

Writes precise prompts for each agent:

• Clear responsibilities
• Workspace paths
• Interface contracts
• Success criteria

Phase 4: Execution

Launches agents in parallel:

• Background execution
• Progress monitoring
• Failure handling
• Automatic retries

Phase 5: Integration

Merges all outputs:

• Conflict resolution
• Code validation
• Test execution
• Build verification

Phase 6: Delivery

Packages final product:

• Generated documentation
• Setup instructions
• Agent attribution
• Deployment options

Agent Roles

Role	When Used
Core Architect	Always (foundation)
Backend Engineer	Web apps, APIs
Frontend Engineer	Apps with UI
Database Engineer	Data-heavy apps
Auth Engineer	User authentication
Integration Engineer	Third-party APIs
Queue Engineer	Background jobs
Security Engineer	Sensitive data
QA Engineer	Complex projects
Documentation Writer	Public projects
Integration Lead	Always (final merge)

Configuration

Edit resources/config.yaml to customize:

defaults:
  max_agents: 12
  min_agents: 3
  timeout_per_agent: 600  # seconds
  retry_failed: true
  max_retries: 2

Examples

Social Media Scheduler (8 agents)

See examples/social-scheduler-swarm.md

Chrome Extension (5 agents)

See examples/chrome-extension-swarm.md

Helper Scripts

Initialize Workspace

./scripts/init-workspace.sh my-project 5 core backend frontend auth integration

Check Status

./scripts/check-status.sh my-project

Integrate Outputs

./scripts/integrate.sh my-project

Output Structure

/workspace/project-name/output/
├── README.md              # Generated documentation
├── package.json           # Dependencies
├── src/                   # Source code
├── tests/                 # Test files
├── docs/                  # Additional docs
├── AGENTS.md             # Agent attribution
└── .gitignore

Metrics Tracked

• Total agents launched
• Time to completion
• Files generated
• Lines of code
• Test coverage
• Build status
• Integration issues

Best Practices

1. Be Specific: The more detail in your request, the better the swarm design
2. Mention Tech Stack: If you prefer React over Vue, say so
3. State Constraints: Timeline, complexity level, must-have features
4. Review the Plan: Agent Architect shows its plan before launching

Limitations

• Maximum 12 agents per swarm
• 10-minute timeout per agent
• Best for well-defined projects
• May need iteration for complex requirements

Contributing

Feel free to:

• Add new agent role templates
• Improve prompts
• Add examples for different project types
• Enhance integration logic

License

MIT License - Use freely, attribution appreciated.

---

The Agent Architect is the conductor. The agents are the orchestra. Together, they build symphonies of code.