🤖 DevFlow-MultiAgent

An End-to-End Autonomous Coding Engine

DevFlow is a state-driven multi-agent framework that takes a natural language requirement and autonomously produces analyzed specs, architecture design, production-grade code, unit/integration tests, and a code review — in one single run.

---

✨ What It Does

Give DevFlow a plain-English prompt like:

"Build a FastAPI REST API for a to-do list with SQLite persistence, CRUD endpoints, and input validation."

And watch five specialized AI agents collaborate to deliver:

Stage	Agent	Output
🔍 Analysis	Requirements Analyst	Structured requirements, constraints, edge cases
🏗️ Design	Solution Architect	Design patterns, component diagram, tech choices
💻 Coding	Senior Developer	Production-ready, documented source code
🧪 Testing	QA Engineer	Unit & integration tests with full coverage
🔎 Review	Code Reviewer	Quality report, approval status, revision notes

If the reviewer requests changes, the pipeline loops back automatically and iterates until the code meets quality standards — up to a configurable maximum.

---

🏛️ Architecture

User Query
    │
    ▼
┌─────────────────────────────────────────────────────────────┐
│                    LangGraph Workflow                        │
│                                                             │
│  ┌──────────┐  ┌──────────┐  ┌──────────┐  ┌──────────┐   │
│  │ Analysis │→ │  Design  │→ │  Coding  │→ │ Testing  │   │
│  │  Agent   │  │  Agent   │  │  Agent   │  │  Agent   │   │
│  └──────────┘  └──────────┘  └──────────┘  └──────────┘   │
│                                                  │          │
│                        ┌─────────────────────────┘         │
│                        ▼                                    │
│                  ┌──────────┐                               │
│                  │  Review  │──── APPROVED ──→ Done ✅      │
│                  │  Agent   │                               │
│                  └──────────┘                               │
│                        │                                    │
│                    REVISE ──────────────────→ Coding ↩      │
└─────────────────────────────────────────────────────────────┘

Key design decisions:

• State-driven routing via LangGraph — each agent reads from and writes to a shared typed state object
• Conditional edges drive the review loop: approved → done, revise → coding
• Prompt engineering is externalized to the /Prompts directory, making agents easy to tune without touching logic
• Mock mode lets you run the full graph offline without any API keys (great for testing the wiring)

---

📁 Project Structure

DevFlow-MultiAgent/
├── Agents/             # Individual agent implementations (Analyst, Architect, Dev, QA, Reviewer)
├── Prompts/            # System & user prompt templates for each agent
├── States/             # Typed state schema for the LangGraph workflow
├── Tools/              # Shared tools available to agents (e.g., Tavily web search)
├── UI/                 # Optional web interface (port 8000 by default)
├── Workflow/
│   └── WorkflowGraph.py  # Graph definition — nodes, edges, routing logic
├── architecture/       # Architecture diagrams and design docs
├── config.py           # Central configuration (API keys, timeouts, paths)
├── invoke_example.py   # CLI entry point with full usage example
└── LICENSE

---

🚀 Quick Start

1. Clone & Install

git clone https://github.com/moonmido/DevFlow-MultiAgent.git
cd DevFlow-MultiAgent
pip install -r requirements.txt

2. Set API Keys

export NVIDIA_API_KEY=nvapi-...     # Required for live LLM inference
export TAVILY_API_KEY=tvly-...      # Required for web search tool

Or create a .env file at the project root:

NVIDIA_API_KEY=nvapi-...
TAVILY_API_KEY=tvly-...

3. Run

# Live mode — real agents, real LLM calls
python invoke_example.py "Build a FastAPI REST API for a to-do list with SQLite"

# Offline / structural test — no API keys needed
MOCK_MODE=1 python invoke_example.py "Build a to-do list API"

---

⚙️ Configuration

All settings live in config.py and can be overridden via environment variables:

Variable	Default	Description
NVIDIA_API_KEY	—	NVIDIA NIM API key
TAVILY_API_KEY	—	Tavily search API key
MOCK_MODE	0	Set to 1 to run offline without API calls
DEVFLOW_MAX_REVIEW_ITERATIONS	2	Max times the review loop can cycle
DEVFLOW_UI_PORT	8000	Port for the optional web UI
DEVFLOW_WORKFLOW_TIMEOUT_SECONDS	900	Global workflow timeout
DEVFLOW_MODEL_TIMEOUT_SECONDS	300	Per-model call timeout
DEVFLOW_OUTPUT_ROOT	(local path)	Where generated files are saved

---

🧩 Tech Stack

• LangGraph — stateful, cyclical multi-agent orchestration
• NVIDIA NIM — fast LLM inference for agent reasoning
• Tavily — real-time web search for the research tool
• Python 3.10+ — async-friendly, type-annotated codebase

---

🛠️ Extending DevFlow

Add a new agent:

1. Create a new file in /Agents/ with your agent class
2. Add its prompt templates to /Prompts/
3. Register it as a node in Workflow/WorkflowGraph.py
4. Wire the edges in the graph builder

Swap the LLM: Update the model name in config.py — the agent layer is model-agnostic as long as the provider is LangChain-compatible.

Add tools: Drop new tool definitions into /Tools/ and bind them to the relevant agent in its constructor.

---

📋 Example Output

============================================================
FINAL STATUS:       completed
REVIEW:             APPROVED
REVISION REASON:    None
ITERATIONS:         1
============================================================

--- ANALYSIS ---
{ "requirements": [...], "constraints": [...], "edge_cases": [...] }

--- DESIGN ---
{ "patterns": ["Repository", "Dependency Injection"], "stack": [...] }

--- CODING ---
{ "files": { "main.py": "...", "models.py": "...", "tests/": "..." } }

--- REVIEW ---
{ "overall": { "approval_status": "APPROVED", "score": 92 } }

---

🤝 Contributing

Contributions are welcome! Feel free to open an issue to discuss ideas or submit a pull request for bug fixes, new agents, or tool integrations.

---

📄 License

This project is licensed under the MIT License — see the LICENSE file for details.

---

Built by Boutmedjet Abd elmoudjib · 2026

"The best code review is the one that happens before you write the code."