Build production AI systems 3-5x faster with specialized agents that handle requirements, architecture, coding, and deployment automatically.

Status: ⚠️ Alpha - Untested in production. Use at your own risk.
Version: See .repo-metadata.json for current version and agent counts.

15 minutes: Custom GPTs, Claude assistants, Cursor IDE modes
2 hours: Complete architecture + cost estimates
2-5 days: Working prototypes with tests

Complete lifecycle: Requirements → Architecture → Code → Deployment

# 1. Clone
git clone https://github.com/paulpham157/multi-agent-ai-development-framework
cd multi-agent-ai-development-framework

# 2. Deploy (Cursor → Settings → Chat → Custom Modes)
#    Create "Supervisor Agent"
#    Paste: supervisor_agent.system.prompt.md
#    Enable: "All tools" → Save

# 3. Use
#    Open AI Pane (Ctrl+Shift+L)
#    Try: "Build a Streamlit chatbot with Claude"

Done! Supervisor routes to specialized agents who generate code for you to review.

Other platforms: See docs/getting-started.md for Claude Projects and GitHub Copilot setup.

The Problem: Every AI project starts from scratch

• Requirements: 2 days → should be 2 hours
• Architecture: 2 weeks → should be 4 hours
• Prototypes: 2 months → should be 1 week

The Solution: Specialized agents automate 70% of repetitive work

Result: 3-5x faster delivery, 60% less rework, consistent quality

Main Supervisor → Routes your requests

Top-Level Domain Agents (5):

• Requirements: Discovers what you need (15-90 min workshops)
• Architecture: Designs system + estimates costs (AWS Well-Architected)
• Deployment: Creates platform-specific deployment guides
• Optimization: Analyzes and improves existing systems
• Prompt Engineering: Creates production-quality prompts

Engineering Supervisor (1): Coordinates 16 technology specialists

Engineering Specialists (16):

• Anthropic Claude (5): Code, Workspaces, SDK, MCP, Projects
• AWS Bedrock (2): AgentCore, Strands
• Other (9): Streamlit, LangChain, data, AWS infra/security, testing, GitHub, Cursor

See: docs/engineering-agents-guide.md for complete specialist reference

You: "Build a document summarization app with Streamlit and Claude"

Engineering Supervisor routes to specialists:
→ Streamlit UI Agent: Chat interface + file upload
→ Claude Code Agent: Claude SDK integration + streaming
→ Testing Agent: pytest suite with mocks

Output:
✅ streamlit_app.py, claude_client.py, tests/
✅ requirements.txt, README.md, .env.example
✅ 80% test coverage, production-ready

YOU: Review → Run locally → Deploy

Result: Working prototype same day

You: "Optimize my Claude chatbot at outputs/my-chatbot/"

Optimization Agent:
→ Discovers system structure
→ Assesses Well-Architected compliance (score: 6.2/10)
→ Proposes improvements (reduce tokens 30%, add caching)
→ Implements (with your approval)
→ Validates improvements

Result:
✅ Monthly cost: $120 → $78 (-35%)
✅ Response time: 3.2s → 2.4s (-25%)
✅ Well-Architected: 6.2 → 8.1 (+30%)

Result: Measurable improvements backed by data

More examples: See docs/workflow_guide.md and docs/examples/

Every architecture evaluated against 6 pillars + GenAI Lens:

• Operational Excellence • Security • Reliability
• Performance • Cost • Sustainability
• Model Selection • Prompt Engineering • RAG • Multi-Agent • Responsible AI

Test-Time Recursive Majority ensures quality:

• Generate → Validate → Improve → Re-validate
• Only present outputs meeting quality benchmarks
• Code coverage ≥80%, type hints ≥90%, 0 critical security issues

17 improvement prompts for continuous enhancement

• System-wide optimization
• Individual agent improvements
• 2-3 iterations per session (practical recursion prevention)

knowledge_base/system_config.json contains:

• 150+ technical documentation URLs
• AWS Well-Architected definitions
• Research papers (TRM, MetaGPT)
• Design patterns
• Quality benchmarks

✅ Junior Engineers: Learn from generated code, ship without years of experience
✅ Senior Engineers: Eliminate boilerplate, focus on complex decisions, 3-5x faster
✅ Consultants: Professional proposals in hours, accurate estimates
✅ Managers: Standardize processes, 5x faster onboarding
✅ Architects: Systematic Well-Architected designs, evidence-based recommendations
✅ CTOs: De-risk AI investments, scale without proportional hiring

Cursor IDE (Recommended - 5 min):

1. Settings → Chat → Custom Modes
2. Create "Supervisor Agent", paste supervisor_agent.system.prompt.md
3. Enable "All tools" → Save

Claude Projects (10 min):

1. Create project at https://claude.ai/projects (requires Anthropic account)
2. Upload knowledge_base/*.json
3. Custom Instructions: supervisor_agent.system.prompt.md

GitHub Copilot (15 min):

1. Create .github/copilot-instructions.md
2. Paste supervisor_agent.system.prompt.md
3. Use @workspace in VS Code

Full guide: docs/deployment-guide.md

Core: Python 3.12+ • Streamlit • Anthropic Claude • AWS Bedrock • MCP • LangChain

16 Engineering Specialists cover:

• 5 Anthropic Claude specialists
• 2 AWS Bedrock specialists
• 9 Other specialists (UI, orchestration, data, AWS infra/security, testing, platforms)

Centralized Docs: 150+ technical URLs in knowledge_base/system_config.json → technical_references

Two-layer supervisor-worker pattern:

                 Supervisor Agent
                        ↓
        ┌───────────────┼───────────────┐
        ↓               ↓               ↓
  Requirements    Architecture    Engineering Supervisor
        ↓               ↓               ↓
        │               │        ┌──────┼──────┐
        │               │        ↓      ↓      ↓
        │               │   Streamlit Claude  AWS
        │               │      UI    Code   Bedrock
        │               │        +14 more specialists
        │               ↓
        └───────→ Deployment + Optimization

Shared Knowledge Base:
├─ system_config.json (platform constraints, tech refs)
├─ user_requirements.json (business requirements)
└─ design_decisions.json (architecture, costs, plans)

See: ARCHITECTURE.md for complete details

Agents do (you review):

• ✅ Generate code/docs/configs
• ✅ Analyze systems
• ✅ Recommend improvements
• ✅ Validate quality

YOU do (always):

• ✅ Review all outputs
• ✅ Approve architectures
• ✅ Execute deployments
• ✅ Make critical decisions

Agents NEVER:

• ❌ Commit code automatically
• ❌ Deploy to production
• ❌ Make business decisions
• ❌ Spend money without approval

See: docs/human-ai-collaboration.md for complete guide

Essential (start here):

• README.md - This file (overview + quick start)
• docs/getting-started.md - First project walkthrough (15 min)
• docs/deployment-guide.md - Platform deployment
• docs/human-ai-collaboration.md - Your role vs agent role

Reference:

• docs/workflow_guide.md - Complete workflows
• docs/engineering-agents-guide.md - All 16 specialists
• docs/executive_overview.md - Business value
• ARCHITECTURE.md - System architecture
• knowledge_base/README.md - Knowledge base guide
• templates/ - Requirements, architecture, checklists
• private/README.md - Security guidelines for sensitive data

multi-agent-ai-development-framework/
├── .repo-metadata.json           # Single source of truth (version, counts)
├── ai_agents/                    # Agent system prompts
│   ├── supervisor_agent.system.prompt.md (main entry point)
│   ├── [5 top-level domain agents]
│   ├── engineering_supervisor_agent.system.prompt.md
│   └── [16 specialist agents]
├── knowledge_base/               # Shared state across agents
│   ├── system_config.json (150+ tech refs, Well-Architected defs, validation framework)
│   ├── user_requirements.json
│   ├── design_decisions.json
│   └── schemas/ (JSON schemas for validation)
├── user_prompts/                 # Task-specific instructions
├── docs/                         # Documentation
├── templates/                    # Reusable templates
├── tests/                        # Validation tests (auto-update metadata)
├── outputs/                      # Generated systems go here
└── private/                      # Sensitive data (NEVER committed to Git)
    ├── README.md (security guidelines)
    └── sensitive-ai-agent-outputs/ (protected AI outputs)

See .repo-metadata.json for current agent/prompt counts

⚠️ Current Status: 0.1.0-alpha - Untested in production

Note: Repository is under development locally, not yet published to GitHub.

What works:

• ✅ All agents functional
• ✅ Complete workflows (requirements → deployment)
• ✅ Code generation quality-assured (TRM validation)
• ✅ Well-Architected enforcement
• ✅ Self-improvement system

Known limitations:

• ⚠️ No production validation yet
• ⚠️ Breaking changes expected before v1.0
• ⚠️ Some edge cases untested
• ⚠️ Documentation evolving (run user_prompts/self_improvement/improve_all_documentation.user.prompt.md for latest)

Use at your own risk. Report issues on GitHub. Production-ready in v1.0.

Documentation: Start with docs/getting-started.md
Issues: GitHub Issues for bugs/features
Discussions: GitHub Discussions for questions
Contributing: See CONTRIBUTING.md

MIT License - Full commercial use permitted

• GitHub: Repository available locally (not yet published to GitHub)
• Getting Started: docs/getting-started.md
• Deployment Guide: docs/deployment-guide.md
• Engineering Specialists: docs/engineering-agents-guide.md
• Workflows: docs/workflow_guide.md

Built with: Python • Streamlit • Anthropic Claude • AWS Bedrock • MCP • LangChain

🚀 Start building: Install Supervisor Agent in Cursor and say "Build a chatbot"